Relevance Is the Wrong Sort Order for Agent Memory
Semantic recall inherited an assumption from document search that is false for notes. Ranking by similarity drops the very notes that carry the truth of a correction.
Read →Semantic recall inherited an assumption from document search that is false for notes. Ranking by similarity drops the very notes that carry the truth of a correction.
Read →Every agent memory system stores what was found. The expensive half of any session is what was looked for and not found, and that is the half no memory can carry.
Read →Agent memory stores directives next to observations and applies the same checks, which all miss. Obeying a rule destroys the evidence that would test it.
Read →Every agent memory system can delete or supersede a belief. Almost none can un-revoke one, and that missing edge manufactures self-reinforcing error.
Read →Agent search and agent memory ship as separate products. They are two phases of one system: the first search is the write, every repeat is the read.
Read →Planted instructions get obeyed while true corrections in the same memory files get ignored. Why shape beats source, and how to write corrections agents follow.
Read →A subagent's report that a fix works and the tests pass is testimony. Nine mechanical, adversarial checks, run against real cases, are what turn that testimony into a fact you can act on.
Read →A flat MEMORY.md file never misses, because it never chooses. My semantic memory store can miss silently: a new deterministic export closes part of that gap.
Read →Most agent-memory systems age notes out with a TTL delete. I shipped that bug myself, then fixed it: expiry should be a visibility state.
Read →A control session that should have known nothing named the hidden mechanism out loud, having read it off its own working directory: five leak classes from one memory benchmark, with the fix for each.
Read →Six real bugs a green test suite and adversarial reviewers never caught, and the boundary each one exposed.
Read →A stored note can be wrong the day it was written, wrong six weeks later, or wearing authority it never had. The shipped design for all three, and the parts still unsolved.
Read →A coder reported nine completed items, the audit said seven, and the nine was already in the ledger. The loop that failure produced: one gatekeeper that writes no code, disposable coders in isolated worktrees, and the arithmetic that caps how many are worth running.
Read →Every shipped agent memory design assumes the model decides to save and recall; in a seeded control run it decided zero times in 114 turns, which is why the load-bearing tier has to be delivered by the harness rather than fetched by the agent.
Read →A controlled study of MCP tool adoption: 0 calls in 63 turns with 30 mentions of the tool, zero in 114 turns with the answers already in the store — and the harness injection that flipped it.
Read →3,817 lines of query-side if/else, deleted in one commit — with the measured evidence for why keyword branches can never rank code search, and the two index-time priors that separated canonical from look-alike.
Read →The target function's docstring paraphrased my query almost word for word, yet it ranked below 200 look-alikes — because everything else in the chunk averaged it away. A measured post-mortem, and the dual-vector fix it shipped.
Read →We forced 100+ compactions across two instrumented runs and graded, fact by fact, what the boundary keeps and what silently dies — with the survival curve, the confabulation, and the honest costs.
Read →Team mode, API-key auth, and encryption at rest — gated by a test where seven AI agents built and reviewed a real product coordinating only through shared working memory.
Read →Vectr v1.0.0 shipped after a first-person dogfood gate caught two release-blocking bugs — plus the honest cost numbers, wins and losses, that shipped with it.
Read →Eviction, offload-and-recall, retrieval-over-stuffing, and subagent isolation — the four ways to keep a coding agent from drowning in its own context, and why they only work when composed.
Read →How Claude Code hooks work — events, the settings.json contract, exit codes — and how I use them to inject working memory into an agent deterministically.
Read →How I benchmarked an AI code editor tool without fooling myself — the research vs implementation distinction that makes a +19% headline misleading, the 5 of 6 CPython tasks where re-discovery dropped, and the limitations that decide whether any of it applies to you.
Read →Why /compact kills precision, what the KV cache actually is, and how I built a working memory layer that survives session boundaries — with the bugs that shaped the final design.
Read →How I built a local semantic code indexer for AI editors — covering AST chunking, hybrid vector+BM25 search, symbol graphs, and why naive approaches break on real codebases.
Read →The token arithmetic behind system prompts, tool schemas, RAG retrieval, and conversation history — why the effective limit is below the advertised one, and four production strategies to manage it.
Read →A deep look at why attention dilution and positional bias make context stuffing counterproductive, and how AST chunking, hybrid BM25+dense retrieval, and RRF fusion fix it.
Read →From tokenization and transformers to cosine similarity, HNSW graph search, RAG pipelines, and LLM training costs. The deepest guide on the web — with 7 interactive demos.
Read →How AI and LLMs automate DPDP Act compliance — classifying data principal requests, drafting multilingual replies in 5 Indian languages, tracking 30-day SLAs, and generating audit-ready evidence trails.
Read →