StrataGate-AgentMemory

A long-running AI agents biggest memory challenge is not storage capacity but accuracy and sufficiency: summaries lose dates and exact wording, similarity search can conflate neighboring events, and frequently retrieved results reinforce themselves into a retrieval loop. StrataGate restructures long-term memory around these four problems. Every memory keeps its L5 source messages permanently; older blocks automatically decay toward lighter index levels. Event cards record both mention time and occurrence time, link back to their source blocks, and track status. After every retrieval, an evidence gate asks whether the current results are genuinely sufficient before the agent answers, directing it to re-search, expand an event, or read the original turns when they are not.

Technically, StrataGate relies on deterministic BM25 plus structured field rankings combined via reciprocal-rank fusion rather than vector retrieval. Events can be projected into a current-state graph describing the present status of people, projects, tools, and places, with every derived fact citing the event that produced it. State changes close or supersede earlier derived facts without rewriting the source event. Reinforcement is deliberately conservative: a retrieval hit does not strengthen a memory; only the items actually used in the final answer gain long-term weight, eliminating self-reinforcing loops.

On LoCoMo conv-26 (419 messages, 152 questions), StrataGate averaged 80.46 percent across ten independent Judge runs versus 63.22 percent for the Mem0 baseline, with the largest gap in temporal questions at +40.27 percentage points. The entire system runs on a local SQLite database with no separate memory server. As a DeepSeek Harness plugin it captures completed turns automatically, builds searchable indexes in the background, and exposes a built-in Memory UI under DSH Settings. It suits developers and advanced users who need cross-session long-term memory that is traceable, verifiable, and local-first.