01
Mine real sessions
The learning workflow examines transcripts for repeated mistakes, corrections, and operational patterns worth carrying forward.
// Operator memory
Agents that learn from their own mistakes
PasClaw can mine past sessions for recurring failures and preserve the resulting lessons as explicit rules the next run can use.
Learn commandSCARS.mdPortable rulesLocal files
02
Preserve failure anchors
SCARS.md records concise reminders tied to past failures so the agent can recognize and avoid familiar traps.
03
Keep humans in control
Proposed rules remain inspectable files that operators can edit, reject, version, and review alongside project code.
04
Export across runtimes
Render one set of operator guidance into the instruction formats used by other agent tools, keeping behavior aligned across environments.
From the documentation
Technical details
Implementation notes drawn from the PasClaw repository documentation.
Local memory layers
The workspace memory directory combines always-on guidance with searchable notes and durable failure reminders.
- MEMORY.md is injected into the system prompt.
- SCARS.md assigns stable section anchors to recurring failures.
- Daily notes and fetched URL bodies remain inspectable files.
Hybrid retrieval and knowledge
FTS5 BM25 provides exact retrieval; optional sqlite-vec, ONNX, and MiniLM embeddings add semantic recall with reciprocal-rank fusion.
- The system falls back to full-text search when vector support is unavailable.
- kb sync chunks operator-curated documents into a separate local corpus.
- memory_search, memory_fetch, session_search, kb_search, and kb_get expose retrieval to the agent.
Source documentation
Common questions
Frequently asked questions
Keep exploring
See how the rest of PasClaw fits together.