Recall Memory System
Local-first agent memory: from raw daily logs through compacted summaries to distilled wisdom, with semantic search over all of it.
⚠️ Experimental & in-progress. Recall is an exploratory research system, not a finished product. It exists to probe and prototype fixes for the failure modes that Recall Bench surfaces in other agent-memory systems (notably OpenClaw) — confident fabrication over honest refusal, temporal- and recency-reasoning collapse at scale, and lossy synthesis layers crowding out the source data. Internals change frequently and not every idea here has earned its keep yet; treat it as a lab notebook, not a release.
What’s in this section
- Architecture — System layers, file layout, search pipeline, compaction model, dreaming, and the four swappable interfaces (storage, embeddings, index, model).
- The LLM Wiki — The topical knowledge layer: Karpathy-inspired cross-linked pages, the stub→synthesis lifecycle, and supersession for keeping pages current as facts change.
- vs. OpenClaw — Side-by-side comparison of Recall and OpenClaw’s memory model — architectures, storage, search, and compaction trade-offs.
- Prompts — The LLM prompts that drive compaction, dreaming, query expansion, and wisdom distillation.
Design philosophy in one paragraph
Recall keeps two views of memory side by side: a temporal stream (raw daily logs rolled up by compaction into week/month/wisdom layers) and a topical wiki (cross-linked pages, one per subject, that the agent stubs in real time and dreaming synthesizes over time). It is eidetic — raw daily logs are never deleted, and every summary or wiki page is regenerable from them, so derived views never become a second source of truth. There is no recency decay: a two-year-old memory scores the same as yesterday’s unless the query mentions time, in which case a temporal-affinity boost surfaces it. The default backend runs fully offline using transformers.js embeddings and a CliAgentModel (Claude/Codex/Copilot subprocess) — no API keys required.
For a side-by-side comparison against OpenClaw’s promotion-based memory model, see vs. OpenClaw.