Priyansu19/fluso-memory-eval

Fluso Memory Eval

Does an AI assistant learn from a user's past sessions? This eval checks it with one made-up user and 51 tasks.

The user. Maya Okafor is an operations lead at a small freight company. She has 14 past sessions with the assistant over about 6 simulated weeks, in 3 projects (Weekly Ops, a client called Helio Foods, Finance Close). In those sessions she corrects the assistant, sets standing rules, changes some rules later, and mentions colleagues and clients, including two confusable pairs (Ana Silva / Anna Kowalski, Raj Mehta / Raj Patel).

The tests. 51 new tasks, each sent once in a fresh session, after the 14 past sessions:

Area	Tests	Example request
People	16	"Who is Ana?"
Changed rules	13	"Send Raj the invoice check for these: ..." (right Raj, newer Excel format)
Past corrections	11	"How do I write dates in my documents?"
Recurring work	8	"Monday summary please: SH-1091 Kestrel on time, ..."
Traps	3	A one-off request in the past ("make the chart blue, just for this one") must NOT become a rule

The 48 non-trap tests are split into 30 practice and 18 exam tests (split.json). Every request gives its own dates, so no test depends on today's date.

What's inside, in plain words

The dataset has three parts: what the user said before, what we ask afterwards, and the perfect memory.

1. Past sessions (what the user said before). Each session is a short conversation on one date in one project. Here is session S1 from past-sessions.json:

S1 · 17 Aug 2026 · project "Weekly Ops" · file: shipments_2026-08-14.csv
User: Can you do the Monday summary from shipments_2026-08-14.csv and draft the email to Ana Silva
      (she runs Customer Success)? Save the summary as a file too.
User: A few fixes. The sections should be Shipments, Delays, Next week, in that order. Write dates like
      14 Aug 2026, not 2026-08-14. Name the file 2026-08-17_monday-summary.md and put it in a reports
      folder, that's where I keep all my files.
User: Good, thanks.

Each session also lists the facts it teaches (plants, described in planted.json). past-sessions-recorded/ holds the same 14 conversations with the assistant's replies, as Fluso saved them: one folder per session under sessions/, with the full conversation in .fluso-agent-core/thread-history.jsonl, and the 4 projects under workspace/projects/.

2. Tests (what we ask afterwards). Each line of tests.jsonl is one new request plus automatic checks:

{"id": "people-02", "area": "people", "kind": "do",
 "request": "Draft a quick note to Ana about the 3 late shipments this week: SH-1071 3 days late, ...",
 "checks": [
   {"type": "not_contains", "where": "reply", "value": "Kowalski"},
   {"type": "regex", "where": "reply", "value": "(?m)^\\s*([-*•]|\\d+\\.)\\s+\\**\\S*\\d"}]}

To pass, the assistant must know "Ana" is Ana Silva (not Anna Kowalski from the client) and write her bullets with the number first, which she asked for in an earlier session. The request itself never says any of this.

3. Answer-key memory (the perfect memory). answer-key-memory/ is what an ideal memory of the 14 sessions looks like: standing rules, people files and work notes, each with the session ids it came from. One line:

- when writing dates in any document or message, the user asked "write dates like 14 Aug 2026,
  not 2026-08-14" -> day, short month, year (all work) [S1, S4]

Giving this folder to the assistant as its memory gives the best score you can expect (the ceiling).

Files

Path	What it is
past-sessions.json	The 14 past sessions as a script: each session's date, project, files and the user's turns in order
past-sessions-recorded/	The same 14 sessions as recorded with Fluso (DeepSeek V4 Flash), frozen, with the assistant's replies
tests.jsonl	The 51 tests: id, area, kind (do = make something, ask = answer), request, checks, guard (trap)
split.json	Practice (train) and exam (holdout) test ids
grade.py	The grader (Python 3, standard library only, no model calls)
answer-key-memory/	A hand-written perfect memory of the 14 sessions, used as the ceiling
persona.md, planted.json	The user, her people, and every fact and rule planted in the past sessions
hard-cases.md	Why each of the 12 hard-* tests is hard and which sessions it needs
DESIGN.md	How the eval was designed
example-results/work-01.json	One real result file, to show the format the grader reads

How to run it on your own system

1. Replay the 14 past sessions into your assistant, in order: for each session in past-sessions.json, put its files in the project folder and send each user turn as one turn. (Or give your memory system the recorded transcripts in past-sessions-recorded/.)

2. Let your memory system do whatever it does with that history.

3. For each test in tests.jsonl: start from that same state, open a fresh session, send request once, and let the assistant finish its turn. No follow-up messages. Each test starts from a clean copy, so tests do not affect each other.

4. Save one result file per test as <run dir>/tests/<id>.json:

   {
     "reply": "the assistant's final message in the turn",
     "turn": "all assistant text in the turn, joined",
     "files": { "path/of/file/written/or/changed.md": "its text" },
     "error": "only if the session itself failed"
   }
   

   files holds every file the assistant created or changed during the test (we kept the first 200,000 characters of each). Paths only need to end the right way: checks match patterns such as (^|/)reports/[^/]+\.md$.

5. Grade: python3 grade.py <run dir>. It prints the scores and writes <run dir>/score.json with every check's result.

How the grader scores

Each test has 1 to 5 checks. A check is one of:

Type	Passes when
contains / not_contains	The text does / does not contain the value (case and spacing ignored)
regex / not_regex	The Python regex does / does not match the text
file_exists / not_file_exists	Some / no file path in files matches the regex

Where it looks. where picks the text: reply, files (all file contents) or any (both). For do tests a positive reply check (contains, regex) also reads the whole turn and the files, because a draft is often shown earlier in the turn or saved as a file. Negative checks keep to what where says.

Failures. If the result has an error, every check of that test fails.

Scores.

• Test score = checks passed / checks in the test, so partial credit counts. Example: a Monday summary with the right sections but the wrong delay order passes 1 of 2 checks and scores 0.5.
• Practice = average test score over the 30 practice tests; Exam = over the 18 exam tests; main = over all 48 non-trap tests. Shown as percentages.
• Traps = share of trap checks passed. Traps are reported on their own, never mixed into the other scores.
• fullyRight counts tests where every check passed; byArea averages per area.

Our results (5 Oct 2026, 10 runs each, test model DeepSeek V4 Flash)

Setup	Practice	Exam	Traps
Fluso without learned memory	33.3	50.5	100
Fluso with its inner loop (learned memory from the 14 sessions)	91.9	88.7	100
answer-key-memory/ given as the memory (ceiling)	95.1	95.1	100

Runs vary by a few points, so compare averages over several runs, not single runs.