# Dynamics-memory

A bounded, self-correcting long-term memory layer for LLM agents that distills noisy, contradictory project logs into a capped memory pool instead of letting a vector store fill up with stale, conflicting copies.

Dynamics-memory is a bounded long-term memory layer for LLM agents. Rather than stuffing an entire log into the context window or storing raw text in a vector database, where redundant and contradicting copies pile up over time, it treats the log as evidence and distills it into memories that must pass a value and confidence gate before entering a capped memory pool.

A background investigator responds to recall misses and triggers such as new decisions or entities by rereading the log and proposing updates, sharing the same ingestion path as normal capture with no special privileges; a value score decays over time while a separate confidence score gates whether a memory is actually served, so a high-value but low-confidence memory stays inert. Contradicting memories are aggregated rather than silently overwritten. It is written in TypeScript, and its mechanisms have been tested in isolation, though an end-to-end run has not yet completed.

Website: https://github.com/1173591564/Dynamics-memory