Analyse-July/05-memory-knowledge.md

> **Hinweis:** Der zugehörige Subagent-Transcript lieferte keine Assistant-Antwort (nur User-Prompt). Dieser Report wurde **direkt aus** `~/.hermes/`, Cursor/Codex-Skills und den fertigen Schwester-Reports rekonstruiert (Stand: Juli 2026).

# Memory & Knowledge Architecture

## Three-tier memory model

Samuel's stack deliberately **splits memory by purpose** instead of one vector DB:

| Tier | Mechanism | What it stores | Authority |
|------|-----------|----------------|-----------|
| **A. Course / project files** | `~/TH-Mannheim/`, project repos, HAI markdown | Durable learning & shipping truth | **Highest** for uni + code |
| **B. Knowledge Graph MCP** | LightRAG-style stores via `knowledge-graph` MCP | Session ingestion, thoughts, cross-project links | **Exploratory / recall** |
| **C. Hermes profile memory** | `memory:` in config + `memories/`, `global_memories/` | Short user profile snippets, nudges | **Low caps** (5k/2.5k chars) |

Plus **intake register** (voice → structured records) and **Plaud/hai-intake MCP** for packet-oriented context.

## Hermes built-in memory (config)

```yaml
memory_enabled: true
user_profile_enabled: true
memory_char_limit: 5000
user_char_limit: 2500
```

Design signal: Hermes memory is a **lightweight personalization layer**, not the cognitive substrate. Heavy context is expected from **files + MCP**.

`global_memories/` holds shared `USER.md` / `MEMORY.md` style manifests — copy-only backups exist, suggesting careful migration history.

## Knowledge Graph MCP (Cursor)

Registered in `~/.cursor/mcp.json` as `knowledge-graph`. Used when Samuel/context is unclear — query before guessing (per cursor workflow rule). Ingestion paths include agent sessions, folders, prompts DB (per kg skill docs).

**Strength:** cross-session **associative recall** without stuffing prompts.  
**Risk:** graph quality depends on ingestion hygiene; stale nodes if not curated.

## Intake & voice memory

`intake-router-hermes` maintains a **routing register** with structured fields (topics, projects, urgency, confidence). `get_context_package` style tools bridge Plaud transcripts to agents.

This is **event-sourced personal memory** — optimized for "what did I say about project X?" rather than embedding search alone.

## Uni learning rule (critical)

From `AGENT_DIRECTORY.md`:

> Source of Truth ist `~/TH-Mannheim/`, **nicht Profilmemory**.

Fach agents must not substitute chat memory for coursework. This prevents **confident hallucination of lecture content** — a common failure mode in student agent setups.

## ADHS profile as meta-memory

`ADHS_LEARNING_PROFILE.md` acts as **procedural memory** — how to teach Samuel — loaded fresh each message. It is not factual storage but strongly shapes retrieval prompts and session shape.

## Gaps vs. GBrain / OneBrain indie pattern

Indie "second brain" setups often run **overnight enrichment** (66 cron jobs in GBrain lore). Samuel has cron/kanban infrastructure but **less autonomous memory consolidation** while asleep.

## Strengths

- Clear **authority hierarchy** (course folders > register > KG > profile memory).
- MCP makes memory **tool-addressable** from Cursor and Hermes.
- Intake pipeline captures **pre-artifact thoughts** that would otherwise die in voice notes.

## Weaknesses

1. **Synchronization** — same fact may exist in KG, register, and a project README with no single merge.
2. **No unified query UI** — Samuel must know *which* tier to query.
3. **Profile memory caps** can truncate nuance unless flushed to files (depends on discipline).
4. **Checkpointing disabled** in HAI flows — memory of *in-progress reasoning* lives in chat logs, not resumable state machines.

## Recommendations

1. Nightly **register → KG** promotion job for high-confidence intake rows only.
2. Tag KG nodes with **authority level** metadata (course / project / thought).
3. Weekly **memory audit** skill: find contradictions between USER.md and project STATE files.
4. Borrow **completion + consolidation cron** from indie setups — one job, not 66.