An agent builder describes a memory layer that prevents false facts by requiring verbatim source quotes and tracking when facts become true, achieving zero false memories across stress tests despite extraction failures.
The bug that started this: an agent confidently used a stale account balance ($12.69) that had been corrected to $30 days earlier. Nothing hallucinated — the number was real, just no longer true. That's the scariest failure class in agent memory, because every retrieval system happily serves it: almost everything treats "most recently written" as "most true." So I built the memory layer for my agents around two rules instead: Every fact tracks WHEN it became true, separately from when it was last mentioned. A newer sentence about an older truth can't overwrite the current value. One deterministic gate decides what gets stored. A fact must carry a verbatim quote from its source document or it's rejected; anything undated or contested is held for human review instead of asserted. The LLM only extracts — plain code does all the judging. Agents downstream can only state facts that made it through the gate. Today I stress-tested the extraction side against 50 dense chunks of my own operational logs (small local models, so the extractor is the weakest link on purpose). It failed four different ways, one per run: first the model invented its own names for facts, then it copied my key definitions back as "values," then it copied the hint terms I gave it, and after upgrading to a bigger model, it started paraphrasing quotes — right values, reconstructed wording, which the exact-substring check refuses. Current fix, running tonight: the model only returns the value plus a short locator phrase, and deterministic code finds the exact span and takes the verbatim quote itself. Stop trusting a model to copy-paste; make code do the verbatim part. Here's the part that made all four failures feel like wins: poison stayed at zero the whole day. Not one false fact entered memory across any run. Every failure mode ended in "refused," never in "stored something wrong." For agents that act on their memory — send things, book things, answer customers — I'll take under-extraction over confident lies every single time. Tonight it replays ~640 chunks of real logs overnight and grades itself against 66 hand-frozen golden facts. Happy to share results (including if it flops) and more detail on the gate design if people are interested. Question for agent builders: how do you handle the stale-re-mention problem — an old value resurfacing in a newer document? Recency-weighted retrieval makes it WORSE, and I rarely see it discussed next to storage/retrieval choices.
The author critiques the idea of agents remembering everything and introduces TrueMemory, a system that converts memories into trait claims with confidence and evidence to better calibrate agent behavior.
A user describes an issue where AI agents confidently retrieve outdated facts from memory layers without flagging changes, and asks the community for solutions to invalidate old memories or track fact freshness.
The author describes building a 'gate' that logs tool calls and verifies that claims in an AI agent's answer correspond to actual log entries, forcing the agent to show its work and reducing hallucinated responses.
A developer tested how small edge models (LFM2.5, Gemma variants) retain a single fact across conversation turns, finding that models often confidently deny knowing information that remains in context, posing a trust issue for agent architectures and suggesting a trade-off between memory and format discipline.
The article reflects on the complexities of AI agent memory beyond simple storage, highlighting challenges such as determining truthfulness, priority changes, distinguishing decisions from noise, and appropriate timing for surfacing context.