Mentedb

MenteDB's argument is that agent memory fails in two places: the agent cannot tell when a fact it holds has been superseded, and it cannot tell when a rule it holds applies. The product answers both with named machinery. Supersession resolves conflicting facts inside the database — you told it Postgres last month and MySQL this month — without spending a model call on every write to decide which one wins. Action rules tag each rule to the action it governs and fetch it when the agent performs that action, rather than keying retrieval on the current conversation, which is why a rule can arrive at the moment it matters instead of scrolling past in a prompt. The storage is its own: a Rust engine with HNSW vector search, a knowledge graph, BM25 + HNSW retrieval fused with RRF, and bi-temporal valid_from/valid_until stamps so you can ask what the system believed on a given date. Around that sit sixteen core systems on the vendor's count — sleeptime enrichment, stream processing, write inference, trajectory tracking, phantom memories for knowledge gaps, interference shielding, pain signals — plus a Memory Router that swaps one base URL so any OpenAI- or Anthropic-compatible client gets persistent memory with your provider key passed through untouched. Where it earns its keep is agent files. The vendor measured the 70 KB ceiling on real public AGENTS.md files from OpenAI, Kiali and Temporal: fully in context, about 17,500 tokens per turn and 80 percent of the file's own rules followed; from memory, 100 percent at 2 to 8 times fewer tokens, at flat cost per turn however large the file grows. That is a vendor measurement across five trials per file, not an independent study — treat it as a strong hypothesis, not a settled result. Weaknesses are real. It is beta and Apache 2.0 with the vendor's own warning that features and the API may change, so integration means adopting the process_turn loop behind an adapter you control rather than pointing an existing vector database at a new endpoint. Self-hosting puts the data, the LLM key and the nightly maintenance schedule on you, and cross-device sync is a cloud-mode feature a single-machine self-hosted setup does not get. The Codex Router is still listed as in development. For teams that want plain vector persistence with no reconciliation, the contradiction machinery is overhead.