knowledge-graph-memory
Knowledge Graph Memory
Bridge agent conversation logs into an Obsidian knowledge graph, giving every new session access to the full reasoning history of all prior sessions.
Quick Start
Install the bridge script
Copy scripts/conversation_history.py into your project's scripts/ directory:
cp scripts/conversation_history.py <your-repo>/scripts/conversation-history.py
chmod +x <your-repo>/scripts/conversation-history.py
Generate conversation docs
python3 scripts/conversation-history.py --force # Full regeneration
python3 scripts/conversation-history.py # Incremental (skip existing)
python3 scripts/conversation-history.py --dry-run # Preview without writing
python3 scripts/conversation-history.py --limit 10 # Last 10 sessions only
Wire into git hooks
Add to .githooks/pre-push (or equivalent):
if [ -f scripts/conversation-history.py ] && command -v python3 >/dev/null 2>&1; then
echo "[pre-push] Updating conversation history..."
python3 scripts/conversation-history.py 2>/dev/null && \
git add docs/conversations/ 2>/dev/null || true
fi
Add smoke check
Add to your smoke gate script:
if [ -f docs/conversations/Conversations.md ]; then
echo "[ok] conversation history MOC present"
else
echo "[warn] docs/conversations/Conversations.md missing"
fi
What It Generates
Conversations.md (MOC)
An index of all sessions grouped by date:
## 2026-03-16
| Session | Branch | Turns | Duration | Topic |
|---------|--------|-------|----------|-------|
| [[session-2026-03-16-21f4eb55]] | `feature/sti-799` | 44 | 4h 15m | Implement AI Core... |
Per-Session Docs
Each session doc contains:
- YAML frontmatter — session_id, branch, tags, wikilinks
- Metadata table — date, duration, turns, tools, attribution stats
- Conversation thread — chronological timeline:
> [!quote]User prompts> [!info]Assistant reasoning (all text blocks, not truncated)> [!example]Tool calls (expandable, with full input details)
- Files touched — all files read/written/modified
- Commits — checkpoint IDs and timestamps
Data Sources
The bridge script reads two sources:
1. .entire/logs/entire.log — Event stream
Session lifecycle events (start/end/turn), checkpoints, attribution stats, phase transitions. Requires Entire to be installed. If not present, script exits gracefully with code 0.
2. ~/.claude/projects/{KEY}/*.jsonl — Transcripts
Full conversation transcripts from Claude Code. The project key is auto-derived from the repo path (slashes replaced with dashes). Each .jsonl file contains every user message, assistant response, tool invocation, and tool result.
Noise Filtering
The script filters out internal noise from raw transcripts:
<task-notification>blocks — internal task system messagestoolUseResultentries — tool→assistant feedback (not real user prompts)<system-reminder>— system injections- Messages < 5 chars — trivial acknowledgements
- Markdown headers inside callouts — converted to bold
- XML/HTML tags inside callouts — stripped
Obsidian Rendering
All content uses Obsidian callout syntax:
- User messages:
> [!quote] **User** (HH:MM) - Assistant reasoning:
> [!info] **Assistant** - Tool calls:
> [!example] Tool Callswith nested>> [!note] **ToolName** — descriptionper tool - Tool details are expanded by default (remove
-for collapsed)
CLAUDE.md Integration
Add to the "Context Acquisition" section:
### Conversation History as Context
Prior sessions are indexed in `docs/conversations/`. Use them to:
- Recall prior decisions before re-solving a problem
- Understand why code looks the way it does
- Resume interrupted work on a branch
- Avoid repeating mistakes from prior sessions
Search: `grep -rl "keyword" docs/conversations/`
Add to "On Session Start" protocol:
7. Scan `docs/conversations/Conversations.md` for prior sessions on current branch
AGENTS.md Integration
Add to working rules:
7. **Check conversation history for prior context** — before starting work on a branch,
scan `docs/conversations/` for prior sessions. Use `grep -rl "keyword" docs/conversations/`
or read `docs/conversations/Conversations.md` for a chronological index.
Graceful Degradation
| Scenario | Behavior |
|---|---|
No .entire/ installed |
Script exits with code 0, skip message |
| No transcripts directory | Script exits with code 0, skip message |
| Different developer machine | Transcripts dir auto-derived from repo path |
| CI (no local sessions) | Smoke warns but doesn't block |
| Pre-push without Entire | Clean skip, exit 0, ` |