OTHER RESEARCH SYSTEMS & COMPONENTS
Dual-ledger agent memory
A portable memory layer that preserves an agent's history across sessions and runtime environments. A summary ledger provides a compact index, while a detail ledger stores structured memories linked to the original conversations. Agents load the index at startup and retrieve further material as needed.
Source code ↗Results & validation
Cross-session recall has been tested in OpenClaw and Codex, and the Kimi adapter has passed interface tests. Retrieval can move from the summary index to detailed memories and then to the sanitised conversation records.
The open-source alpha supports idempotent writes, concurrent-write locking, source tracking, audit and retraction, with a rebuildable SQLite full-text search index. Full end-to-end testing with the Kimi model has not yet been completed because of usage limits.
Implementation
Session archiving
Available conversations are sanitised and archived. Knowledge, procedures and events are extracted as structured records, each retaining a reference to its location in the original conversation.
Ledger maintenance
Accepted memory records are stored in the detail ledger. A small selection of lasting value is indexed in the summary ledger, with full content retrieved when needed.
Layered recall
At startup, the agent reads the summary index. It uses SQLite full-text search to retrieve details from the detail ledger, then consults the full sanitised conversation if the structured records are insufficient.