Evolving Agents Labs

Evolving Agents Labs/Experiments/evolving-memory

evolving-memory

What if an agent's memory consolidated itself the way sleep consolidates yours?

Results Apr 2026
Many irregular traces funnelling into a few consolidated ones

Traces in, structure out

An agent generates enormous volumes of execution trace and almost none of it is worth keeping verbatim. The engine chunks traces, connects related ones, and curates what survives — a consolidation pass, not an append-only log.

What comes out has properties you can state and test: repeated experience raises confidence, failures extract constraints rather than just being recorded, and one domain's lessons do not bleed into another's.

There is also an instruction set and a small VM — traces become programs with opcodes, versioned through a registry — plus SQLite and FAISS storage, three LLM provider backends, and a 14-route FastAPI server.

What's proven

7,449 lines across nine subpackages with 182 test functions in 14 files. The interesting ones are behavioural rather than structural: test_failure_extracts_constraints, test_repeated_experience_increases_confidence, test_semantic_isolation_between_domains, test_multi_domain_agent_lifecycle. Those are claims about what the engine does, checked in code.

The honest status

This engine has no consumers. Nothing in the organization imports it — skillos_robot decoupled from it in April, and every other mention is prose. It is a well-tested, framework-agnostic, genuinely reusable component that nobody is currently plugging in.

That is a statement about attention, not quality. But you should know it before adopting it: it has been cold since April, and if you hit a bug you are likely the first.

← token-trie