Lightweight agent memory systems
Historical coverage in the frozen legacy collection. New doc-grounded or
code-grounded analyses use the main analysis method
and publish under kb/agentic-systems/.
Paper-, article-, product-, or spec-grounded systems that are useful for the agent-memory-systems landscape but have no inspectable implementation that supports a code-grounded review in ../reviews/.
These historical agent-memory-system-review notes carry
source-tier: doc-grounded. Their legacy comparison contract is retained in
the review type.
- AgeMem — RL-trained memory-management policy covered through source ingest and analysis notes.
- Fintool — production AI agent for professional investors; commercial-scale evidence for filesystem-first, covered from a founder's practitioner report with no public source.
- Incremental Self-Improvement — Schmidhuber's reward-gated self-modification paradigm for learning learning strategies; useful lineage for policy-level promotion, but lightweight coverage and not a modern KB system.
- Mnemosyne / IsaacCLupus mnemosyn spec — spec-first local semantic memory OS with schema/prototype evidence but no current reference implementation.
- Sig — macOS work-memory app covered from public release docs, local-file claims, and product README material rather than inspectable app source.
- Trajectory-Informed Memory Generation — trajectory-to-tip learning pipeline covered through source ingest.
TODO: OpenClaw-RL now has a reachable repository (Gen-Verse/OpenClaw-RL) and should get a repo-backed review rather than a lightweight note.
Complete file listing (generated at build time)