Self Imprection

README

# self-imprection

Standalone OpenClaw plugin concept for:

- session reflection
- self-improvement / governance
- selective long-term storage through RAG
- Markdown-first audit trails and human review

This repository started from design-first docs and now contains a minimal working plugin skeleton.

The goal is to keep it separate from `memory-lancedb-revised` so the memory engine and the governance/reflection workflow can evolve independently.

## Intended Position

`memory-lancedb-revised` remains the memory runtime:

- hybrid retrieval
- scopes
- rerank
- auto-capture / auto-recall
- compact-v2 pre-ingest

`self-imprection` is intended to become the workflow/governance runtime:

- post-session reflection
- durable decision extraction
- lessons / pitfalls extraction
- self-improvement logs
- promotion pipeline for stable rules and reusable skills
- selective write-back of durable items into long-term RAG memory

## Design Docs

- [Architecture](./docs/architecture.md)
- [Storage Tiering](./docs/storage-tiering.md)

## Status

Minimal working skeleton:

- session-boundary reflection hooks
- structured reflection artifact generation
- Markdown reflection writing
- storage-tier candidate extraction
- RAG write-back bridge with `queue-file` and `direct-lancedb` modes

Recommended rollout:

1. enable plugin with `queue-file` mode first
2. inspect reflection artifacts and queued RAG candidates
3. switch to `direct-lancedb` only after reviewing candidate quality and scope policy

## Current Local Rollout

The current OpenClaw config has already been wired to load `self-imprection` with:

- `reflection.enabled = true`
- `ragWriteback.enabled = true`
- `ragWriteback.mode = "queue-file"`

This means:

- reflections will be generated at session boundaries
- Markdown artifacts will be written
- RAG candidates will be queued to JSONL files instead of being written directly into LanceDB

Recommended next live check:

1. trigger `/new` or `/reset`
2. inspect generated reflection Markdown under the workspace memory tree
3. inspect queued JSONL candidates under the configured `memory/rag-writeback/` directory
4. only then consider switching `ragWriteback.mode` to `direct-lancedb`