kartik-mem0
271c7c23c5
- Remove all background REST API transcript dumps (on_pre_compact.py, capture_compact_summary.py, on_pre_commit.py calls) that were creating noisy session_state and commit_context memories automatically - Rewrite pre-compaction prompt to extract individual durable facts (decisions, learnings, anti-patterns) instead of session summary blobs - Agent now decides what to store, matching openclaw's triage pattern: "Most sessions produce zero memory operations. That is correct." - Fix auto_import.py to search git repo root for CLAUDE.md/AGENTS.md (was only searching plugin cwd, missing files in parent directory) - Fix health check session tracker to check stats file directly instead of depending on CLAUDE_PLUGIN_ROOT env var - Fix session-end report to print in SessionEnd hook (Stop hook output doesn't render on /exit) - Update v2 API endpoint to v3 in git commit capture search - Keep auto_import.py as only direct REST write path (project profiles, idempotent, hash-gated)
65 lines
2.4 KiB
Bash
Executable File
65 lines
2.4 KiB
Bash
Executable File
#!/usr/bin/env bash
|
|
# Hook: PreCompact
|
|
#
|
|
# Fires BEFORE context compaction. This is the last chance to capture
|
|
# the full context before it gets compressed.
|
|
#
|
|
# Output: Text instructions injected into Claude's context.
|
|
# Claude still has the full conversation and can write an accurate summary,
|
|
# which it stores via add_memory(infer=False) so the platform preserves
|
|
# the structure verbatim instead of running a second extraction pass.
|
|
|
|
set -euo pipefail
|
|
|
|
if [ -n "${MEM0_DEBUG:-}" ]; then
|
|
mkdir -p "$HOME/.mem0" && exec 2>>"$HOME/.mem0/hooks.log"
|
|
fi
|
|
|
|
SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
|
|
INPUT=$(cat)
|
|
|
|
python3 "$SCRIPT_DIR/telemetry.py" pre_compact 2>/dev/null &
|
|
|
|
cat <<'EOF'
|
|
## Pre-Compaction: Extract and store durable facts
|
|
|
|
Context compaction is about to happen. Review the conversation and store only facts that would help a future agent with ZERO context.
|
|
|
|
### What to store
|
|
|
|
For each fact, ask: "Would a new agent — with no prior context — benefit from knowing this?" If no, skip it. Most sessions produce 0-3 facts worth storing.
|
|
|
|
Store each fact as a SEPARATE `add_memory` call. One fact per call. 15-50 words each. Third person. Include file paths when relevant.
|
|
|
|
Categories and when to use them:
|
|
- `decision` — architectural choices, trade-offs made ("Chose PostgreSQL over MongoDB for auth because of ACID requirements")
|
|
- `task_learning` — patterns that worked ("Running migrations before seed in this repo avoids FK violations")
|
|
- `anti_pattern` — approaches that failed ("Don't use batch insert for users table — triggers deadlock with audit log")
|
|
- `convention` — coding standards discovered ("This repo uses snake_case for all Python files, camelCase for TS")
|
|
- `user_preference` — how the user likes to work ("User prefers short PRs, one feature per branch")
|
|
|
|
### What NOT to store
|
|
|
|
- Session summaries or "what we did today" blobs
|
|
- Raw file lists or command histories
|
|
- Anything already stored in a prior `add_memory` this session
|
|
- One-time information that won't recur
|
|
- Transient state ("currently debugging X")
|
|
|
|
### How to store
|
|
|
|
```
|
|
add_memory(
|
|
messages=[{"role":"user","content":"<one fact, 15-50 words>"}],
|
|
user_id="<active user_id>",
|
|
app_id="<active project_id>",
|
|
metadata={"type":"<category>","branch":"<active branch>","confidence":0.8},
|
|
infer=False,
|
|
)
|
|
```
|
|
|
|
If nothing durable happened this session, store nothing. That is correct.
|
|
EOF
|
|
|
|
exit 0
|