ai agent memory infrastructure

What you could build here

Supply is crowded, with 88 launches in 28 days and only 8 above 10 upvotes. The recurring complaints are memory bloat, stale or wrong memories, and agents ignoring their own instruction files, and most of these were raised in only one thread each, so treat them as leads to verify.

1. Memory compaction with history kept. Automatically merges and prunes stored agent memories while keeping a record of what was removed. Aimed at developers whose agent notes grow until retrieval gets expensive. Raised in 2 threads: agents can't decide what to keep, creating bloat, and a request for cleanup without losing history.
2. Approval queue for agent-written memories. Holds memories an agent proposes until a human approves, edits, or rejects them. Aimed at teams whose agent notes are too verbose or unreliable to trust unreviewed. Raised in 2 threads: agent memories need editorial review and are overly verbose, and auto-extracted memory is no better than hand-written docs without curation.
3. Contradiction detection and belief revision. Flags memories that conflict with newer facts and lets the agent or user correct early wrong assumptions. Aimed at developers whose agents keep acting on outdated beliefs. Raised in 4 threads across contradiction handling, belief revision, and entity review requests. None of the example launches mention it.

What people complain about in these tools

• Memory bloat, irrelevant context, and poor forgetting make retrieval costly and noisy (9 mentions)
• Stale, wrong, or hallucinated memories keep agents acting on bad information (5 mentions)
• Agents ignore config files, markdown instructions, or skills as context grows (3 mentions)
• Latency and cross session sharing are still weak points (2 mentions)

Tools people mention in this niche: Mem0, Zep, Cognee, llm-wiki, OpenClaw

Written from this niche's launches and Hacker News comment threads. A shortlist to investigate. Check your own idea → Updated Oct 25.