A local-first memory platform for AI agents: durable enough to inspect, retrieve, reinforce, and export.
Agent memory is often treated as an opaque vector store. Fuli explores a more inspectable foundation: canonical records, append-only events, retrieval filters, lifecycle controls, and exportable data.
SQLite is the canonical store. An append-only event log feeds a mutable episode projection, while hybrid vector and FTS5 retrieval can fall back to lexical or substring search. Namespaces, metadata filters, scoring, reinforcement, and consolidation queues sit above that base.
agent event → canonical store → episode projection
↘ retrieve → filter → reinforce
↘ export / diagnostics / metrics| Evidence | What the repository shows |
|---|---|
| Storage | SQLite WAL is the canonical local store with migrations and append-only events. |
| Retrieval | Hybrid vector and FTS5 retrieval includes deterministic fallback paths. |
| Interfaces | MCP and HTTP API surfaces are documented alongside standardized errors. |
| Verification | Offline retrieval benchmarking, diagnostics, Prometheus metrics, and JSONL export/import are included. |
Agents need memory that can be inspected and migrated without depending on a hosted provider.
A useful system still answers when embeddings or a model endpoint are unavailable.
Scoring, reinforcement, soft deletion, and consolidation shape what the agent is allowed to remember.
Export, diagnostics, MCP, and HTTP make memory a component rather than a hidden dependency.