After 25 weeks: four graphs, no LLM in promotion path, and 3 ways agent memory broke
Conscious_Detail_128 · reddit · 2026-09-13
A developer shares a full post-mortem of their agent memory system after 25 weeks in production.
Core design
- Tiered memory: working → episodic → semantic, with nightly consolidation.
- No LLM anywhere in the promotion path: promotion runs on the system's own signals (recurrence, access patterns), not on an extraction prompt — the author argues outsourced extraction makes memory a third-party summary of your agent. Cost: dumber logic that misses things.
- Four graphs instead of one typed graph: an untyped associative graph (edges strengthen on co-access, decay when unused), plus separate graphs for evidence chains, causal structure, and temporal snapshots — association, provenance, causation and change-over-time are distinct relationships that create ambiguity when collapsed.
- Hypothesis-first retrieval: generate what an answer would look like and search on that; keyword/similarity only as fallback. This exposed a vector store that was broken two ways at once: nearly empty and returning NaN-scored fake matches.
- Contradictions as records, not gates: rule-based detection writes conflicts as records mapped to questions and a daily density metric; spikes are healthy.
Three defects found in audit: semantic promotion reads a fixed tail of the episodic ledger that's always too recent, so nothing gets promoted; the noise filter excludes exactly the pattern of the most frequent writer, which can never be promoted; consolidation prunes weak links that index rebuild recreates the same night (link log is 200× memory size).
The lesson that generalizes: the associative graph settled into attractor basins — max-weight cliques from repetitive templated content — after which retrieval returned basin nodes regardless of query. Hebbian reinforcement plus any chatty logger eventually eats your retrieval.
More from coding & agent
- astra is great at CAD itself, but telling it what to build remains the hard part — yacineMTB · 2026-09-13
- Yacine designs drone entirely from phone; custom NPU kernels max out chip without heatsink — yacineMTB · 2026-09-13
- GPT-Live-1 first impressions: most natural voice model yet, but instruction following is unreliable — kolchinski · 2026-09-13
- Open-source 'no-ai-slop' removes 20+ AI writing clichés, hits 8.7k stars — tom_doerr · 2026-09-13
- Devin SWE-2 first impressions: fast and cheap, but shorter endurance than Codex — CtrlAltDwayne · 2026-09-13
- Roboflow ships auto-label with GPT-6 Astra, making it the new default model — Roger_M_Taylor · 2026-09-13