Agent Memory Ledger

Agent Memory Ledger bridge

Use this skill when the user asks to remember, archive, or recall durable facts, procedures, decisions, or events across OpenClaw sessions.

Boundary

The standalone ledger directory is the canonical memory write target. This bridge must not write MEMORY.md, memory/, OpenClaw Memory SQLite, or another platform-owned store. When the operator explicitly configures external-memory mode, a separate provider such as MemPalace may receive only sanitized evidence or derived objects after successful AML archival; it must retain AML archive/object provenance.

Install the core package into a workspace-local virtual environment once:

python skills/agent-memory-ledger/scripts/install.py \
  --source /path/to/agent_memory_ledger-0.1.1-py3-none-any.whl

Then use these local paths:

PY="$PWD/.agent-memory-ledger-venv/bin/python"
BRIDGE="$PWD/skills/agent-memory-ledger/scripts/openclaw_bridge.py"
LEDGER="$PWD/.portable-memory"

Archive the current session

Resolve the current agent id and exact session key from runtime context. Then:

Archive only when the user affirmatively asks to archive the current session with the configured trigger or an equivalent such as "archive", "archive this session", “归档本会话”, or “按 AML 归档”. Questions, explanations, quotations, hypotheticals, and negations about archiving do not trigger the workflow. First extract durable knowledge, procedure, and event candidates into a JSON file as described in docs/AGENT_INSTRUCTIONS.md. Use tags for finer subtypes and canonical kind names only. Set promote: true only for the small set of stable, high-value items that should be discoverable at startup.

"$PY" "$BRIDGE" archive \
  --agent <agent-id> \
  --session-key <exact-session-key> \
  --ledger "$LEDGER" \
  --candidate-file /path/to/memory-candidates.json

Every accepted active candidate is represented in the dark ledger. The bridge uses candidate promote values to build the bright ledger. Add --promote-markers only when the user explicitly supplies stable lines such as Fact:, Rule:, Procedure:, Decision:, 事实：, 规则：, or 流程：. Never edit evidence, objects, ledgers, SQLite, or the journal directly. Run validation after archival and report the archive, object, and promoted IDs.

The bridge exports the OpenClaw trajectory, keeps only visible user/assistant text, converts it to the generic session schema, and calls the standalone core. Never archive hidden reasoning, tool arguments, tool outputs, system prompts, or provider metadata.

Recall

"$PY" "$BRIDGE" recall \
  --ledger "$LEDGER" \
  --query "<user question>" \
  --top-k 6

Answer only from returned active objects. Include relevant object_id values so the result is auditable. Do not silently fall back to OpenClaw memory search. An explicitly configured external provider may be used for deeper recall, but its result must retain or resolve to AML archive/object provenance.

Validation

"$PY" "$BRIDGE" validate --ledger "$LEDGER"

Completion requires ok: true, at least one evidence manifest, structured objects, both ledgers, and successful cross-process recall.