funes now recalls in more than English
dacorvo
• id stringlengths 4 4 | kind stringclasses 3
values | question stringlengths 28 159 | gold listlengths 1 1 | quote stringlengths 18 300 |
|---|---|---|---|---|
q001 | verdict | How much did the denormal fix speed up the blas backend on the Intel box? | [
{
"session_id": "53a0f4d7-ad90-41b7-ac06-a0edef9f69d3",
"from": 0,
"to": 39
}
] | Stopped — not running anything further.
Your two runs reproduce it, decisively. Baseline = `26d6150` (bench workload, no fix), fixed = `48c486a`/`pr-148` (+ the 46-line `FtzDazGuard`):
| workload | embed base → fix | rerank base → fix | speedup | subnormal softmax weights* |
|---|---|---|---|---|
|
q002 | verdict | On the Intel box, how did the two candidate fixes for the denormal stall compare? | [
{
"session_id": "53a0f4d7-ad90-41b7-ac06-a0edef9f69d3",
"from": 43,
"to": 83
}
] | **The simpler fix matches the guard exactly.**
| 16×mixed (ragged batch) | embed ms | rerank ms |
|---|---|---|
| base — no fix | 5880.0 | 11186.1 |
| `FtzDazGuard` (+46 lines) | 1150.3 — **5.11×** | 2534.8 — **4.41×** |
| `vexp` select-zero (+15 lines) | 1137.0 — **5.17×** | 2524.9 — **4.43×** |
|
q003 | verdict | On a Mac, why was ORT slower than torch for the reranker? | [
{
"session_id": "6282863b-f46b-45c2-b2ad-cd53cb5db658",
"from": 414,
"to": 454
}
] | Definitive — and it confirms the mechanism cleanly (this microbench is contention-robust: both backends, same shapes, same window):
| GEMM (reranker's linears) | torch / Accelerate | ORT / MLAS | MLAS slowdown |
|---|---|---|---|
| [7680×768]×[768×768] | 1163 GF/s | 233 GF/s | **4.98×** |
| [7680×7 |
q004 | verdict | How did torch and ORT compare for the reranker on a Mac across input sizes? | [
{
"session_id": "6282863b-f46b-45c2-b2ad-cd53cb5db658",
"from": 377,
"to": 417
}
] | This is the payoff — the small-input sweep **corrects the whole picture**. Same window, back-to-back, quiet machine:
| input | ort (funes) | torch CPU | ratio |
|---|---|---|---|
| [1, 32] | 19.1 ms | 23.5 ms | **0.81×** (ort faster) |
| [1, 128] | 57.7 ms | 61.1 ms | 0.94× |
| [8, 128] | 428.6 ms |
q005 | verdict | How much faster was the blas backend's 256×mixed workload on a Mac with PR #168? | [
{
"session_id": "2026-09-23T07-16-55-506Z_01a0cd1f-e852-7b29-b33f-9a0b06c57161",
"from": 238,
"to": 278
}
] | Now we have clear, reliable numbers:
**256×mixed workload (1 iteration, 0 warmups):**
Baseline (main):
- embed: 29762.9 ms (~30 seconds)
- rerank: 61658.5 ms (~62 seconds)
- Total real time: ~2m49s (because embed + rerank + overhead)
PR #168 (length-sorted grouping):
- embed: 2644.6 ms (~2.6 seco |
q006 | verdict | On a Mac, what did PR #168 do for the blas backend's 16×mixed rerank? | [
{
"session_id": "2026-09-23T07-16-55-506Z_01a0cd1f-e852-7b29-b33f-9a0b06c57161",
"from": 116,
"to": 156
}
] | The median results from runs 1-3 show:
- 16×short: embed +1%, rerank +7% (slight slowdown for PR)
- 30×~500tok: embed -4%, rerank +3% (mixed)
- 16×mixed: embed +8%, rerank -44% (huge rerank speedup!)
- 32×real: embed 1290.6, rerank 3209.5 (no baseline available, but we can infer)
The `16×mixed` rer |
q007 | verdict | What did the evaluation of semantic query expansion conclude? | [
{
"session_id": "291ec080-9a3d-48b7-aa8a-5f85c6b78cbb",
"from": 74,
"to": 114
}
] | All measurements are in. **F4 is a no-go: semantic expansion is strictly dominated by the plain `--candidates` knob at every matched budget.**
## The matched-budget verdict
Same box, same store, same gold; pool sizes are mean unique turns:
| pool size | expansion (all-gold / any-gold) | fused-N c |
q008 | verdict | In the recall-features error analysis, where did the gold turns end up in recall's results? | [
{
"session_id": "291ec080-9a3d-48b7-aa8a-5f85c6b78cbb",
"from": 193,
"to": 233
}
] | ## Step 1 results — the error analysis
**Where the 886 gold turns actually are** (turn-level, best chunk rank):
| | c30 | c100 |
|---|---|---|
| in top-10 | 36% | **36%** (unchanged) |
| in pool, below top-10 | 7% | 22% |
| never retrieved | 57% | 41% |
Everything the deeper pool adds lands below |
q009 | verdict | Before the Hub cache existed, how much data did a recall over a Hub memory download? | [
{
"session_id": "2a73dd48-c0db-444c-b6d6-ce50620d4471",
"from": 59,
"to": 99
}
] | Both runs confirm it: **44.0 MB / 47.4 MB** transferred for two cold recalls of the *same* query, with **no drop on the repeat** — so there's no cache today (the feature's motivation), and a single recall reads ~20% of the 216 MB store via byte-ranges. The 278s-vs-783s spread is exactly the "unpredi |
q010 | verdict | What did enabling LTO do to the size of funes's macOS release binary? | [
{
"session_id": "614d266e-0ffa-48d7-b3b3-da7bc5b79491",
"from": 13,
"to": 53
}
] | Verified — runs fine (`funes 0.10.1+dev`, all subcommands intact) and the symbol table is gone (only undefined externs remain).
## Result
**221.7 MiB → 118.6 MiB — 103 MiB saved (46.5%)**, build ~23 min.
Where it came from (segment before → after):
| Segment | Before | After | Saved by |
|---|-- |
q011 | verdict | On a Mac, how did the LTO variants compare for funes's release build? | [
{
"session_id": "614d266e-0ffa-48d7-b3b3-da7bc5b79491",
"from": 41,
"to": 81
}
] | Both built cold. Here's the comparison:
| Profile | Size | vs 222 MiB baseline | Cold build |
|---|---|---|---|
| no LTO (original) | 221.7 MiB | — | — |
| **fat** LTO (committed) | **118.6 MiB** | −46.5% | 22m 57s |
| **thin** LTO | **131.8 MiB** | −40.6% | 14m 21s |
**Thin costs +13.2 MiB (~11% |
q012 | verdict | In the handoff-vs-recall benchmark, what did the /compact arm's summary cost to produce? | [
{
"session_id": "42011d6b-a393-41ff-94bd-4483f7636f32",
"from": 131,
"to": 171
}
] | All three steps are done locally; nothing is uploaded yet.
**1. Compaction re-produced** (one turn, `--resume $FORK_ID --fork-session … --print "/compact"`, model `claude-opus-4-8`, cwd a worktree of `funes-bench-clone` at `BASE_COMMIT`):
| | |
|---|---|
| cost | **$3.67** (8,215 output tokens, 11 |
q013 | verdict | How did the agent demo over huggingface/funes-memory go? | [
{
"session_id": "0e535b7b-ce81-4f8b-894e-62a1ed07a09e",
"from": 74,
"to": 114
}
] | **Decisive.** From a neutral cwd, with no "use recall" instruction, claude called recall **3 times** unprompted — and returned the *counterintuitive* finding it would otherwise get backwards:
> *No — the funes project's past sessions concluded it does **not** help… the "F4 semantic-expansion gate"… |
q015 | verdict | Why were some merged local-plus-remote recalls slower? | [
{
"session_id": "401ba38b-2848-41a8-b273-62670b185a67",
"from": 373,
"to": 413
}
] | The remaining recall cost comes from reranking: when the remote search underfills candidates (24 of 30), local rows fill in and add ~0.6s from the extra 6 reranked docs, while queries where both sides return 15 already match in timing (5.10s vs 5.11s). |
q017 | verdict | How long does rebuilding a large funes memory's indexes take on a 64-core Linux machine? | [
{
"session_id": "d6275c46-408e-4836-9460-708908361082",
"from": 133,
"to": 173
}
] | The fresh baseline is set: rebuilding both indexes over 492k rows took 18s on this 64-core box, uploading about 120MB of index files. Next I'll add remote read timing while the stale variants run. |
q018 | verdict | How fast is first-time indexing of Claude sessions on a Mac? | [
{
"session_id": "79055e24-0af8-46cd-9b58-02f35fd08dd3",
"from": 24,
"to": 64
}
] | You're right, and the measurement proves it. I was hand-waving. Here are real numbers on your Mac.
**Measured throughput:** ~9.9k chars/sec end-to-end (from the `--limit 5` run: 3.72M chars in 371s; the marginal rate excluding model load is ~9.9k/s). This is a worst-case sample — the most-recent se |
q019 | verdict | What did caching do for unit_keys()? | [
{
"session_id": "98360d42-9595-461c-9fbd-e12134b726ae",
"from": 155,
"to": 195
}
] | Cache is in `4ee678b`. Measured again on your stores:
| | before | after |
|---|---|---|
| `unit_keys()` on `.claude/projects` | ~5ms | ~0.10ms |
| `units()` on `.claude/projects` | 7–8ms | ~2.7ms |
`units()` got faster too — it shares the walk, so only the 672 stats and the sort remain. On a netw |
q020 | verdict | What difference did the push reindex fix make on the Open-SWE traces memory? | [
{
"session_id": "6ab35a1e-58fb-42fc-8d58-f6c3a36aa53a",
"from": 209,
"to": 249
}
] | The A/B test is done, and the result is stronger than "faster" — **the fix is the difference between ~80 seconds and never finishing at all.**
## Setup
Duplicated `dacorvo/funes-nvidia-Open-SWE-Traces` (61 GB, 17,098,459 chunks, with a real accumulated backlog of **6,942,054 unindexed rows** acros |
q021 | investigation | How serious was issue #56 about truncated recall hits? | [
{
"session_id": "d165a9a2-2859-476f-8ca2-961c8626cd4b",
"from": 0,
"to": 39
}
] | **Verdict: it's bad — for the MCP/agent path this is likely the highest-impact open correctness issue, because clipping is the common case, not the edge case.** The issue is still live: [render.rs:32](src/render.rs#L32) caps each hit's preview at 400 chars, and MCP `recall` ([mcp.rs:92](src/mcp.rs#L |
q022 | verdict | What did the review find about the update check added to MCP status? | [
{
"session_id": "d442497c-cc51-41da-bb3b-2dea80daa800",
"from": 118,
"to": 158
}
] | The review confirmed one real issue (medium): I put an **unconditional, uncached network round-trip into the MCP `status` tool** — every agent call now risks up to a 5s stall and turns a previously offline-fast local operation into a network-dependent one. The reviewer correctly notes it's not a cor |
q023 | verdict | What did testing funes add's handling of a missing memory turn up? | [
{
"session_id": "adc42fa9-8b40-4f84-9a4e-035d2188c15f",
"from": 868,
"to": 908
}
] | Results:
- **parse_confirm** test passes.
- **Non-interactive + token → stays local** (`['funes','mcp']`), and returns fast — confirming it short-circuits *before* `whoami` (no network, no prompt). ✓
- Explicit-missing store non-interactive: it *bound* the nonexistent store silently rather than erro |
q024 | verdict | How fast was Claude's Stop hook indexing once it ran automatically? | [
{
"session_id": "b5cecf7e-fd8b-4f14-85df-1c35abf0cdac",
"from": 109,
"to": 149
}
] | Steady state confirmed — 12:53:59→12:54:19, **6 new chunks in 0.4s**, `index: ok`. This session's chunk count is climbing cleanly turn over turn: 277 → 284 → 290, each fire appending just its delta. Working exactly as designed. 👍 |
q025 | verdict | What did the first automatic indexing by Claude's Stop hook pick up? | [
{
"session_id": "b5cecf7e-fd8b-4f14-85df-1c35abf0cdac",
"from": 105,
"to": 145
}
] | Confirmed working end-to-end. Two Stop fires logged since the cutover, both `index: ok`, lock cleaned up:
- **12:46 (first fire, one-time catch-up):** 5 sessions / **717 chunks** — this session plus the subagent traces that had piled up (the workflow's `agent-*` transcripts: 84 + 43 + 196 + 117, an |
q026 | verdict | On the Linux box, what did buffer reuse do for the blas backend's speed? | [
{
"session_id": "6be804aa-8022-47ab-9acc-024bd3d225a8",
"from": 347,
"to": 387
}
] | Buffer reuse now beats the env hack with zero environment dependence (3.19s / 1.62s). Still 1.4× behind ort — re-profiling to find the next bottleneck: |
q027 | verdict | On the Linux box, how did the blas backend compare with ort? | [
{
"session_id": "6be804aa-8022-47ab-9acc-024bd3d225a8",
"from": 927,
"to": 967
}
] | Yes — on this Linux box, blas rerank is slower, and consistently so: ~18% at the recall worst case (2.59s vs 2.19s) and ~1.8× on short batches (487ms vs 270ms). Embedding is near parity (within 3–6%). That's the honest trade as measured, and it's worth restating what it does and doesn't cost:
- **I |
q028 | verdict | When funes push was made safe for two machines pushing to the same memory at once, did that slow down the push round-trip test? | [
{
"session_id": "98d5ea58-667e-4ee5-bbbf-4485d7bb20a1",
"from": 282,
"to": 322
}
] | Round-trip green with the pinned path: **84.4s vs 83.5s** before it — the extra `head_oid` + id scan per attempt costs nothing measurable (the id read is warm-cached by sha).
Both gated tests now pass on the current tree:
- `push_round_trip` — one racer publishes, the other hits the lock and steps |
q029 | verdict | In the shallow-indexing review, how much overhead did the change add to an index run with nothing to do? | [
{
"session_id": "cd7dca1f-743e-4c91-a14d-1c56f1d673d9",
"from": 589,
"to": 629
}
] | I went through the review and changed one thing.
- **Concern 1, the full vector-column scan:** measured rather than estimated, the extra cost is small. On a clone of your real memory (214k chunks), an idle budgeted run takes 0.162 s against main's 0.098 s. That 64 ms is the scan plus the text-index |
q030 | verdict | Why did a private dataset nobody fetched show so many downloads? | [
{
"session_id": "019f855a-62c0-7e81-a287-774bc03f100a",
"from": 647,
"to": 687
}
] | That explains it.
The failed range-read attempt likely issued many `HEAD` and ranged `GET` requests. Hugging Face counts those as downloads even when only a few bytes are fetched or the operation ultimately fails. Requests are deduplicated only per IP, repository, and five-minute window.
So `1,598 |
q031 | decision | Which agents was funes meant to support, based on their audience? | [
{
"session_id": "0298d3b2-eb4d-41e0-98e0-abd5e4801adc",
"from": 144,
"to": 184
}
] | focus on hermes only. Opencode is really low audience in the ML community, just as pi. Hermes is 10x these agents in terms of usage, and Claude Codex are 20x and 10x Hermes respectively |
q032 | decision | What was decided about funes's built-in hello corpus? | [
{
"session_id": "0298d3b2-eb4d-41e0-98e0-abd5e4801adc",
"from": 655,
"to": 695
}
] | I think we can drop the hello database now, since the user is guided by funes add anyway |
q033 | decision | What was the view on funes index publishing to the Hub on its own? | [
{
"session_id": "1a0ff281-dcee-4422-8862-3a8dfe68b7da",
"from": 440,
"to": 480
}
] | anyway, the feature is dangerous, because you might push inadvertently to the wrong remote, thinking you are only indexing |
q034 | decision | Where should a store lock be implemented in funes? | [
{
"session_id": "7a7321e6-1347-46c9-8c12-be0b667680e9",
"from": 75,
"to": 115
}
] | If a lock is really required, it does not belong to a script, but to funes itself |
q035 | decision | What was decided about the scope of a harness integration bundle? | [
{
"session_id": "67013480-25ae-4860-bf1f-591c6b3c4857",
"from": 352,
"to": 392
}
] | yes, i don;t want any bundle versb except add remove. It is up to the bundle to figure out how to make its turn files available. This has to be clearly explained that funes will only index what'sin its harness spool |
q036 | decision | Where should the migration of legacy funes installs be handled? | [
{
"session_id": "91c9a93a-3ded-4ae7-bf3e-aec047279919",
"from": 212,
"to": 252
}
] | because it is a no go from my side: if you need something to be fixed, do it in the setup for each harness |
q037 | decision | What happened to the secrets fix proposed on PR #137? | [
{
"session_id": "5fb86142-cc70-4026-af54-68f5368961bb",
"from": 17,
"to": 57
}
] | I don;t do half-fixes. I prfer the current error that blocks indexing than the one that will let the secrets be leaked silently |
q038 | decision | What was the plan for issue #179? | [
{
"session_id": "dae3bece-7afd-4900-acfe-3dc18c889dc1",
"from": 31,
"to": 71
}
] | I certainly won't add any workaround. What could be done is bump rmcp
then reexamine our options |
q039 | investigation | Which change made the rows_with_ids filter slow? | [
{
"session_id": "970dec84-a4cb-4db4-9b32-44a00908e328",
"from": 88,
"to": 128
}
] | Tracked it down. The filter itself is old, but the change that made it reach its worst case is **PR #123 "Remove curation"**, merged 2026-08-28 — 11 days ago.
## The lineage
| date | commit | what it did to the query |
|---|---|---|
| 2026-06-23 | [cbc4cac](https://github.com/huggingface/funes/com |
q040 | investigation | Why was funes still pinned to lance 7, and how far behind the latest major was that? | [
{
"session_id": "72d9aaba-702f-44d1-8869-a7fa3917bdad",
"from": 25,
"to": 65
}
] | **Release dates** (stable majors, from the lance repo):
| version | released | notes |
|---|---|---|
| 6.0.0 | 2026-05-11 | |
| **7.0.0** | **2026-05-27** | our pin |
| 8.0.0 | 2026-07-01 | |
| 9.0.0 | 2026-07-24 | |
| 10.0.0 | 2026-08-07 | |
| 11.0.0 | 2026-08-30 | current latest |
| 12.0.0-beta.9 |
q041 | investigation | How old were funes's pinned embedder and reranker models in July 2026? | [
{
"session_id": "019f8d9b-5bb5-7ed1-8967-864a59ca9f87",
"from": 451,
"to": 491
}
] | As of July 23, 2026, both models are nearly three years old:
| Role | Model | First published | Age |
|---|---|---:|---:|
| Embedder | `BAAI/bge-small-en-v1.5` | Sep 12, 2023 | ~2 years, 10 months |
| Reranker | `BAAI/bge-reranker-base` | Sep 11, 2023 | ~2 years, 10 months |
Sources: [embedder his |
q042 | investigation | Does funes's CI run all the integration tests? | [
{
"session_id": "ebcf7ae4-d61b-4335-86f7-0c17a8958c54",
"from": 695,
"to": 735
}
] | Right — and CI proves it: PR #128 is **all green** while carrying tests that fail locally.
`.github/workflows/ci.yml` never runs the whole suite. The hermetic job runs `cargo test --lib --profile ci` — unit tests only — and the integration job names its targets one by one. Diffing what's on disk ag |
q043 | investigation | Where did the MCP protocol version pin in funes come from, and was there a reason for it? | [
{
"session_id": "dae3bece-7afd-4900-acfe-3dc18c889dc1",
"from": 215,
"to": 255
}
] | `361cc3f` (2026-06-18), the initial commit. No rationale in the message, and it was a deliberate downgrade — rmcp's `ProtocolVersion::default()` was already `LATEST` (2025-11-25) then, so someone typed that line on purpose.
Most likely copied from rmcp's example servers, which used that constant. N |
q044 | investigation | What caused the CI failure in pi's startup test? | [
{
"session_id": "91c9a93a-3ded-4ae7-bf3e-aec047279919",
"from": 920,
"to": 960
}
] | Found that the pi startup test used a faulty fake, masking a real crash: an unhandled EPIPE on the child's stdin when funes exits early. I'll add a stdin error handler in the extension and fix the test with a proper echo fake plus a dying-funes case. |
q045 | decision | Which memory should the `→ get` hint under each recall hit name? | [
{
"session_id": "32cde944-9842-41bd-b097-1b3894a5900e",
"from": 336,
"to": 376
}
] | sorry, but if you print a hint, then print the store the recall just hit, period |
q046 | decision | What was decided about recording the harness in funes memories? | [
{
"session_id": "78d4da76-56ec-4f75-a98f-bacb21b3bddc",
"from": 129,
"to": 169
}
] | Not crystal clear I concede, but yes, I want the harness ot be included in the memories |
q047 | decision | Which lance operation was chosen to add the harness column to existing memories? | [
{
"session_id": "78d4da76-56ec-4f75-a98f-bacb21b3bddc",
"from": 157,
"to": 197
}
] | I prefer add_columns |
q048 | decision | What was required of the turns-file indexing path when converters moved into bundles? | [
{
"session_id": "0d4eeff6-83b9-48cc-93d0-44c8d3354fea",
"from": 291,
"to": 331
}
] | i donlt want to tie this to B. A+C must be equivalent to main without B |
q049 | decision | How should partially indexed spool turn files be handled? | [
{
"session_id": "0d4eeff6-83b9-48cc-93d0-44c8d3354fea",
"from": 322,
"to": 362
}
] | this is far easier to solve (did not read your prose): I think we can just leave the turn files that have not been completed yet, and just do a sweep once we're done with the passed turn file |
q050 | decision | What is the recommended way to refresh an agent integration that is already installed? | [
{
"session_id": "ef57c2b4-4a4c-470a-83e4-c25ac84cc871",
"from": 221,
"to": 261
}
] | ok, on existing installs, I want a note somewhere the preferred way is always to funes remove then funes add |
q051 | decision | Where does per-agent automation belong? | [
{
"session_id": "bdfccb0c-89ed-4bee-9fc7-f98a859600ee",
"from": 219,
"to": 259
}
] | automation has nothing to do inside funes |
q052 | decision | What was decided about the --update flag of funes add? | [
{
"session_id": "6b8338c4-a701-4823-8a79-116b42bf8c77",
"from": 120,
"to": 160
}
] | delete the flag then |
q053 | decision | What was the view on having --porcelain next to --format for recall's human output? | [
{
"session_id": "91dda941-fff1-4f38-bfb3-868d96a8f2c5",
"from": 225,
"to": 265
}
] | I don't buy the dual flags --format human == --porcelain. This is awkward and error prone + harder to maintain |
q054 | investigation | What did the experiment comparing ways of running funes ask show? | [
{
"session_id": "c634b7f3-7f70-4ee9-83d4-e4736bb28a6b",
"from": 222,
"to": 262
}
] | The experiment is done — two stress-phrased questions, three arms each, run from a neutral cwd. Both questions were chosen so the top-8 recall would struggle, and the recall logs confirm it (near-zero scores). Here's what happened:
| | Q1 — answer *was* reachable from top-8 | Q2 — answer *not* in t |
q055 | investigation | Could raw Codex traces be fetched from the Hub? | [
{
"session_id": "78d4da76-56ec-4f75-a98f-bacb21b3bddc",
"from": 77,
"to": 117
}
] | Yes — Codex raw traces are fetchable. The earlier failure was the fix: byte-range requests return the LFS/Xet *redirect* body, so `-L` on the whole file (they're small, 35 KB–17 MB) is the way. I pulled two and mapped the full native shape, including tool use — so this closes the Codex half of the P |
q056 | investigation | Why did enter do nothing in the fzf picker? | [
{
"session_id": "8e87000b-7a29-498c-b30b-bd0fa6f29301",
"from": 132,
"to": 172
}
] | Done, and all checks pass (both clippy configs, 17/17 test targets green, release rebuilt — so the `~/.local/bin/funes` symlink already serves the fix).
**What was happening:** enter was genuinely doing nothing — the bind was `cat {4} | ${PAGER:-less -R}`, fzf runs binds through `$SHELL -c`, and zs |
q057 | investigation | What did the review of PR #183 find? | [
{
"session_id": "01a0d871-53c2-7900-8559-c040b569630d",
"from": 0,
"to": 28
}
] | I found a concrete budget regression: rejected files count toward neither batch limit, so a malformed backlog can be processed in full before the next 60-second budget check, holding the memory lock throughout.
The redaction mapping looks sound, including a live check with the same secret repeated |
q058 | investigation | When does the session sketch cache get cleared? | [
{
"session_id": "772c4a20-4161-4bfb-bad7-fe1ecd96518b",
"from": 34,
"to": 74
}
] | Never — nothing deletes it. There's no TTL, no size cap, and no command that clears it.
What exists is *invalidation*, not clearing: [session_sketch.rs:355-372](src/session_sketch.rs#L355-L372) rejects an entry whose schema version, selector version, memory label, session id, source fingerprint, em |
q059 | investigation | Why did the funes 0.11.0 release build fail? | [
{
"session_id": "91dda941-fff1-4f38-bfb3-868d96a8f2c5",
"from": 739,
"to": 779
}
] | Root cause confirmed from the log. The kill lands **six minutes into compiling `funes v0.11.0` itself** — the final crate, which is exactly where fat LTO with `codegen-units = 1` merges the bitcode of the *entire* dependency graph (lance, arrow, ort, tokio…) into one giant module:
```
21:22:28 Com |
q060 | investigation | Why does the LongMemEval store's name end in _s? | [
{
"session_id": "c9a1043b-ed39-4e04-98a9-6196e778310a",
"from": 99,
"to": 139
}
] | Because the store holds only the **`_s` variant's haystack**. LongMemEval ships three files, and each defines a different corpus:
- `longmemeval_s` — ~50 sessions per question; deduped, that's the 19,195 sessions / 311k chunks being indexed now.
- `longmemeval_m` — ~500 sessions per question (2.6 G |
q061 | verdict | Did the fix for issue #110 work on its test case? | [
{
"session_id": "06299b3e-e644-4626-b195-d50706a0b7d0",
"from": 66,
"to": 106
}
] | Verified end-to-end: 37.1MB / 24 findings → 0.3MB / **0 findings**. Now the full check: |
q062 | verdict | How did the local backfill of the repo column go? | [
{
"session_id": "a48aee15-9cca-40c5-ab74-351053bed2f5",
"from": 17,
"to": 57
}
] | Backfill succeeded — **46/46 sessions resolved** (no unresolved gaps), `repo` column added to 8,760 chunks. Let me verify it landed and is idempotent, and confirm the new schema. |
q063 | verdict | Did the shared chunking step make funes index --check slower? | [
{
"session_id": "0d4eeff6-83b9-48cc-93d0-44c8d3354fea",
"from": 157,
"to": 197
}
] | Shared step added, and all affected tests pass including redaction; the check test slowed from 2.9s to 7.5s due to per-unit scanner spawn overhead, matching real run costs. Now re-running the full suite on the refactored code. |
q064 | verdict | What came out of the isolated install and upgrade scenarios run before the 1.4.0 release? | [
{
"session_id": "bdfccb0c-89ed-4bee-9fc7-f98a859600ee",
"from": 123,
"to": 163
}
] | All eight scenarios passed with zero failed checks and every `funes add` exiting 0. Now I'll pull specific evidence from the re-run upgrade logs — the stale-install notice, drain counts, and pi's legacy-path handling — before writing up the results. |
q065 | verdict | What stack size did the push round-trip test need in CI, on the branch where it overflowed? | [
{
"session_id": "2026-09-28T08-55-40-395Z_01a0e73a-1c6b-7317-bb18-d009d37ef24c",
"from": 65,
"to": 105
}
] | 4MB suffices, 2MB (default) doesn't. So the test thread needs between 2MB and 4MB of stack in the `ci` profile. The default test-thread stack in libtest is actually 8MB?? Hmm, no — apparently the default was fine on main and breaks on the branch, so the branch added on the order of ~hundreds of KB t |
q066 | investigation | When Hub writes were added to funes in June, what protected against two writers at once? | [
{
"session_id": "3eff638e-e6c4-4225-bd57-c20ded9af143",
"from": 1175,
"to": 1215
}
] | **No.** Verified from primary source across all three layers:
- **HF Hub protocol** — writes are the git-style commit API (`preupload` → NDJSON `commit` → LFS/XET). The *only* compare-and-swap is the repo-level **`parentCommit` SHA**; per-file entries carry no ETag, `If-Match`, or `If-None-Match`. |
q067 | investigation | Why did funes's first release workflow fail for x86_64 Linux? | [
{
"session_id": "3eff638e-e6c4-4225-bd57-c20ded9af143",
"from": 515,
"to": 555
}
] | The re-run got further but **failed at a different spot** — and it's a real one, not a flake:
**Build Linux (x86_64) failed at the link step** (compile succeeded):
```
rust-lld: error: undefined symbol: __isoc23_strtoll / __isoc23_strtoull / __isoc23_strtol
>>> referenced by ... libort_sys (onnxr |
q068 | investigation | What storage would supporting hermes and opencode sessions cost funes? | [
{
"session_id": "3eff638e-e6c4-4225-bd57-c20ded9af143",
"from": 429,
"to": 469
}
] | Now I can address both properly.
## Storage cost of the export
Measured from the real datasets — two numbers matter: the **export size** (the materialized text) and the **indexable fraction** (what funes keeps):
| agent | native store | export size | indexable text (what funes keeps) |
|---|---|- |
q069 | investigation | Which long-term memory benchmark was chosen to evaluate funes, and why? | [
{
"session_id": "3eff638e-e6c4-4225-bd57-c20ded9af143",
"from": 1139,
"to": 1179
}
] | Strong signal — this is genuinely worth doing, and one fact makes it a clean fit: **LongMemEval explicitly separates indexing / retrieval / reading, and ships an official retrieval-only scorer.** So funes can be evaluated on exactly what it is (a retriever) without ever generating an answer.
## The |
q070 | decision | How does funes's approach to memory differ from hermes's memory providers? | [
{
"session_id": "3eff638e-e6c4-4225-bd57-c20ded9af143",
"from": 1076,
"to": 1116
}
] | this is not what funes was made for. funes bets on the MODEL to revisit its memory, it does not proactively inject anything |
q071 | decision | Why doesn't funes summarize sessions into facts? | [
{
"session_id": "3eff638e-e6c4-4225-bd57-c20ded9af143",
"from": 1013,
"to": 1053
}
] | One thing that also drives my intuition towards NOT synthetizing the memoreis is that i trust future models to be a lot better at doing it, possibly crossing references also |
q072 | investigation | Why did funes ask get the cost comparison with compaction wrong, and was the grounding branch needed? | [
{
"session_id": "02f1c668-964e-4920-88f6-d41e6371488d",
"from": 371,
"to": 411
}
] | Your test holds, and I just reproduced it independently: I ran the rephrased question through a chunks-only prompt — exactly what `main` builds — and it answered correctly *and* attributed `$0.19–0.25` to compaction's task turn, calling out that it ignores the one-time summarizing cost. That is prec |
q073 | investigation | What did the A/B of funes ask prompt variants on the compaction cost question show? | [
{
"session_id": "02f1c668-964e-4920-88f6-d41e6371488d",
"from": 140,
"to": 180
}
] | Done. Six answers across five prompt variants, one query, against `huggingface/funes-memory`. The public results settle the ground truth first:
**Ground truth** (`dacorvo/funes-handoff-recall-benchmark/results/README.md`): yes — C recall **101k** vs E compact **778k** weighted tokens per successful |
q074 | decision | What was the view on adding session details to the prompt funes ask builds? | [
{
"session_id": "02f1c668-964e-4920-88f6-d41e6371488d",
"from": 145,
"to": 185
}
] | the identity lines are of no use to the agent being queried |
q075 | verdict | Is the Claude run by funes ask isolated from the user's own configuration? | [
{
"session_id": "02f1c668-964e-4920-88f6-d41e6371488d",
"from": 78,
"to": 118
}
] | The child does load `/home/ubuntu/.claude/CLAUDE.md` — your global instructions ride along, so "the passages are your complete context" isn't strictly true. Now the new query. |
q076 | verdict | Can the agent behind funes ask still run tools? | [
{
"session_id": "02f1c668-964e-4920-88f6-d41e6371488d",
"from": 75,
"to": 115
}
] | Confirmed empirically: under those flags the child attempted exactly one `tool_use`, and it came back `is_error:true` with permission text — attempted, blocked.
Your follow-up is the sharper point though: it doesn't know funes, unless the cwd tells it. Checking that while I set up the new query. |
q077 | decision | In the shallow-index work, what was decided about the BTree index on chunk ids? | [
{
"session_id": "39e85166-d113-4cd6-8c3e-d95217b4e204",
"from": 75,
"to": 115
}
] | [tool_result AskUserQuestion] Your questions have been answered: "B1 pairs delta `optimize_indices` with a BTree scalar index on `id`. The BTree was measured necessary on lance 7 (Auto mode); on lance 11, `MergeInsertWriteMode::RewriteColumns` patches vectors in place without it, and the index only |
q078 | verdict | Did the rmcp upgrade in PR #184 break anything in the MCP server? | [
{
"session_id": "01a0d91c-8cc1-7792-8560-3380ebaca371",
"from": 0,
"to": 25
}
] | No new finding from the standalone lifecycle probe.
Using rmcp 3.4.1 with Funes’s `ServerConfig` construction and generated tool handler:
- Conformant `server/discover`, followed by `tools/list` and `tools/call`, succeeds; process exits **0** after stdin closes.
- Legacy `initialize` with `2025-11 |
q079 | investigation | When the issue about Chinese sessions came in, what was the history of choosing the embedding model in funes? | [
{
"session_id": "5f22631b-72fc-4715-8609-483e8cc54359",
"from": 44,
"to": 84
}
] | Checked. Here's what I found.
## The history: it was never per-model flexible
The model has been hardcoded since the very first commit (`361cc3f`): `const MODEL: &str = "BAAI/bge-small-en-v1.5"; const DIM: i32 = 384;`. There's no `--embedding-model` flag anywhere in the history.
What you actually |
q080 | investigation | How well could funes handle sessions mostly in Chinese before the multilingual work? | [
{
"session_id": "5f22631b-72fc-4715-8609-483e8cc54359",
"from": 44,
"to": 84
}
] | Checked. Here's what I found.
## The history: it was never per-model flexible
The model has been hardcoded since the very first commit (`361cc3f`): `const MODEL: &str = "BAAI/bge-small-en-v1.5"; const DIM: i32 = 384;`. There's no `--embedding-model` flag anywhere in the history.
What you actually |
q081 | decision | With more than one embedding model, what was ruled about sharing a memory between contributors? | [
{
"session_id": "5f22631b-72fc-4715-8609-483e8cc54359",
"from": 66,
"to": 106
}
] | Yes, it is pretty much settled by memories can only be contributed to when the local and remote models are identical. But this is not the main use case. For a personal memory it really does not matter. For a project memory it forces one model for all contributors, which is a project decision, so it |
q082 | investigation | Why did every push from the Linux box fail in early September? | [
{
"session_id": "5f22631b-72fc-4715-8609-483e8cc54359",
"from": 200,
"to": 240
}
] | Root cause, fully evidenced. Not the name, not the store, not push.
**The two "data files" in this host's HF cache contain an S3 error page:**
```
chunks.lance/data/1100010110111101…lance hub 178,041 B local 111 B
chunks.lance/data/1111110001100000…lance hub 178,105 B local 111 B
<?xml ve |
q083 | decision | How should funes deal with a corrupted Hub cache? | [
{
"session_id": "5f22631b-72fc-4715-8609-483e8cc54359",
"from": 201,
"to": 241
}
] | I'd better include something to recover from such situations, so that funes fixes it |
q084 | verdict | What happened when the PR #145 cache repair ran on the Linux box? | [
{
"session_id": "5f22631b-72fc-4715-8609-483e8cc54359",
"from": 265,
"to": 305
}
] | Done and verified against the real poisoned cache.
## What changed
**`fetch_store.rs`** — two things, logically separable for commits:
1. *Correctness*: `serve_from_file` now resolves ranges with object_store's own `GetRange::as_range(total)`, which rejects a request the file cannot span instead |
q085 | decision | What concern came up when integrations became setup scripts under ~/.funes/agents? | [
{
"session_id": "300bf97d-0991-4428-913e-3e4c7043b167",
"from": 370,
"to": 410
}
] | unless I miss something, this is a very nice backdoor you introduced here, allowing a rogue user to bypass the offiical integration script and have his installed instead, if it manages to put them on the user disk. What mitigations do we have ? Also, when fetching/using, we should ask a trust confir |
q086 | investigation | Why did the Claude Code integration's setup script change how it writes its hooks? | [
{
"session_id": "300bf97d-0991-4428-913e-3e4c7043b167",
"from": 482,
"to": 522
}
] | Good instinct, and it was worse than it looked. python3 is out, committed as `000df76`.
**Why it was a real risk.** On a fresh Mac, `/usr/bin/python3` exists on PATH but is a Command Line Tools stub that fails until Xcode CLT is installed. And node is no longer implied either, since Claude Code shi |
q087 | decision | How should the Codex integration be installed? | [
{
"session_id": "300bf97d-0991-4428-913e-3e4c7043b167",
"from": 498,
"to": 538
}
] | this is the last resort. I still want the "deterministic" funes add path |
q088 | investigation | Before the integration registry, how did funes add pi install the pi extension? | [
{
"session_id": "300bf97d-0991-4428-913e-3e4c7043b167",
"from": 286,
"to": 326
}
] | No, not the install script. The funes binary itself carried the bytes.
`src/agents/pi.rs` had two `include_str!` constants, `index.ts` and `package.json`, read from `integrations/pi/` at compile time and baked into the executable. `funes add pi` then called an `extract` function that wrote those tw |
q089 | verdict | How was the PR #176 fix for a first funes add on a fresh machine verified? | [
{
"session_id": "dcd18946-731f-467d-997f-572852df1233",
"from": 86,
"to": 126
}
] | The first-index gap is closed on `agent_registry`, in two commits on top of the drain:
- **704b4d9 refactor(index): seed a spool by its path alone.** The seed's `Harness` argument was redundant. A spool's turns name their own harness and the turns-file source ignores the override, so `run_index_see |
q090 | verdict | What could the CLI tests for a drop-in integration not cover, and why? | [
{
"session_id": "dcd18946-731f-467d-997f-572852df1233",
"from": 113,
"to": 153
}
] | C3's code gap is closed in commit 56046a6. `funes add` now resolves an integration that exists only under `~/.funes/agents`, with no checkout, published copy, or network. The fallback that `remove` already had moved into the resolution both verbs share. The copy runs unvouched: funes says why it cou |
q091 | decision | Can funes add's trust confirmation be skipped? | [
{
"session_id": "dcd18946-731f-467d-997f-572852df1233",
"from": 116,
"to": 156
}
] | i still don't want any add to be possible without the user being pro;pted |
q092 | decision | How should funes handle sub-agent sessions once agents are no longer known by name? | [
{
"session_id": "dcd18946-731f-467d-997f-572852df1233",
"from": 152,
"to": 192
}
] | this is purely a claude thing, so they shiuld be no specifics about subagents into funes |
q093 | investigation | How did issue #109's claims hold up against the code, and where was its fix? | [
{
"session_id": "d1133fad-7ad1-4f8b-9949-70e79b4c37b1",
"from": 0,
"to": 37
}
] | You remembered right that there was pushback on the issue, but wrong about the merge: **`improve_empty_host_onboarding` is not merged.** `git cherry origin/main improve_empty_host_onboarding` marks its one commit `+` — no equivalent upstream. Issue #109 is still open. The branch survived my cleanup |
q094 | investigation | What was still open on issue #109 once its index half was done? | [
{
"session_id": "d1133fad-7ad1-4f8b-9949-70e79b4c37b1",
"from": 46,
"to": 86
}
] | Here's the full picture. Branch is `eb29f91`, 1 ahead / 0 behind `origin/main` — the index half is done.
## What your earlier decisions actually settled
Worth separating, because it changes what's still on the table:
- **The wrong-memory guard stays.** `must_confirm` is a deliberate design; not t |
q095 | decision | What was decided about the issue #109 documentation? | [
{
"session_id": "d1133fad-7ad1-4f8b-9949-70e79b4c37b1",
"from": 189,
"to": 229
}
] | We still don't need to have it written in the doc what every single action does in detaisl. This is a neverending source of changes |
q096 | investigation | What could be salvaged from the curate gating fixture branch, checked against the issue #109 work? | [
{
"session_id": "d1133fad-7ad1-4f8b-9949-70e79b4c37b1",
"from": 229,
"to": 269
}
] | Nothing is salvageable — both commits were superseded, one of them by you a couple of hours ago.
**`bf45b38 test(curate): put the gating fixture in the project's scope`** — dead. It fixes `tests/curate_gating.rs`, which no longer exists: `9c9cf9d refactor!: remove the project-memory concept` (2026- |
q097 | verdict | When the review of PR #129 resumed, what did the test suite show? | [
{
"session_id": "2026-09-01T09-36-46-330Z_01a05c54-08fa-700a-a770-a7eb5bb11126",
"from": 7,
"to": 47
}
] | Everything passes now:
- `cargo fmt --check` ✅
- `cargo clippy --all-targets` ✅
- `cargo clippy --all-targets --no-default-features --features onnx` ✅
- `cargo test` ✅ (all tests passed)
We also fixed two integration tests that were broken by the branch changes:
1. `tests/add_codex_hooks.rs` — skil |
q098 | investigation | Was the fix for issue #190 already on main when the issue was filed? | [
{
"session_id": "2026-09-28T06-32-47-123Z_01a0e6b7-4b12-7545-b89b-8ef6e0f36dc6",
"from": 20,
"to": 60
}
] | Now I see the full picture clearly:
**On `main`** (commit `d98b8e0`), the `build_indexes` function at `src/memory/dataset.rs` still does:
- `create_index(..., replace=true)` for FTS
- `create_index(..., replace=true)` for IVF_PQ
This is the **old code that rebuilds everything every time**.
**On t |
q099 | investigation | For issue #141 about git worktrees, what does funes record about a session's repository? | [
{
"session_id": "01a07b13-b7f8-7970-901b-164c778f3a24",
"from": 0,
"to": 20
}
] | Yes—“indexing” is the precise term here, and that changes the interpretation.
Funes is **not blind to the repository**. When indexing a transcript, it:
1. Reads the session’s recorded `cwd`.
2. Runs `git -C <cwd> remote -v`.
3. Normalizes the remotes to `owner/name`.
4. Stores that repository iden |
q100 | decision | What was the answer on issue #141 about git worktrees? | [
{
"session_id": "01a07b13-b7f8-7970-901b-164c778f3a24",
"from": 0,
"to": 21
}
] | My answer: "Hi @evgenyfadeev, thank you for your interest and feedback: much appreciated.
About your question: at indexing time funes recognizes that the sessions belong to the same git repository, but not that they were conducted against different branches or commits of that repository. However, if |
q101 | decision | What was the view on the docs written for the budgeted indexing step of PR #196? | [
{
"session_id": "01a0e718-b78d-7020-8493-5f34725518d6",
"from": 0,
"to": 13
}
] | about that lasty commit. I feel we're writing too many details in the docs, leading to a lot of changes down the line every time the code changes |
q102 | verdict | How did the JavaScript pi converter compare with funes's parser on a real pi history? | [
{
"session_id": "agent-aa7a014bc0360a164",
"from": 160,
"to": 200
}
] | Report delivered. Summary of what was done:
**Parity achieved.** A rewritten `poc/converters/pi/convert.mjs` (217 lines, node/JS, zero dependencies) produces byte-identical output to the in-tree parser on the fixture and on all 24 real pi sessions — identical per-file turn/chunk counts, 656 turns / |
End of preview. Expand in Data Studio
funes recall benchmark
Labeled questions for scoring funes recall. Each set is labeled over the funes memories it names,
and its directory holds the questions, how they were written, how they are scored and every run's
receipt.
| set | questions | memories | gold |
|---|---|---|---|
| funes-memory | English | huggingface/funes-memory |
a turn and the 40 before it |
| sharegpt | each in Chinese and English | huggingface/funes-sharegpt-zh, huggingface/funes-sharegpt-zh-e5 (Chinese questions), huggingface/funes-sharegpt-en (English questions) |
a conversation |
A question's gold is one or more spans of its set's memories, each a session_id with from and
to turns or the whole session when they are absent, so a set scores only against those memories.
score_recall (benchmark/score_recall.rs in the
funes repository) asks a set's questions at funes recall's
defaults and prints the receipt: a question's rank is the position of its first hit in a gold span,
and the receipt reports hit@1, hit@5, found@8, the mean reciprocal rank and the median latency. Each
set's README says how to run it.
- Downloads last month
- 338