https://swapnanilsaha.com/blog/ Long-form engineering deep dives and research on retrieval, agent memory, and context engineering. Every number traces back to an instrumented run. en-us Wed, 30 Sep 2026 09:00:00 +0530 swapnanilsaha26@gmail.com (Swapnanil Saha) https://swapnanilsaha.com/blog/when-a-memory-should-become-a-check/ https://swapnanilsaha.com/blog/when-a-memory-should-become-a-check/ Wed, 30 Sep 2026 09:00:00 +0530 Advice needs the agent to believe it; a check does not. But an enforced rule can never be tested again. Which memories should be checks, and what a check must keep. https://swapnanilsaha.com/blog/a-lesson-is-a-failure-plus-a-change/ https://swapnanilsaha.com/blog/a-lesson-is-a-failure-plus-a-change/ Sun, 27 Sep 2026 09:00:00 +0530 Agent memory saves what the model chooses to save. Learning runs on surprise: a failure, a change, then a pass. Without the change it is a retry, not a lesson. https://swapnanilsaha.com/blog/identical-chunks-are-not-duplicates/ https://swapnanilsaha.com/blog/identical-chunks-are-not-duplicates/ Thu, 24 Sep 2026 09:00:00 +0530 Copied scripts must collapse into one search result; generated accessors must not. Their text has the same shape, so near-duplicate dedup needs signals outside it. https://swapnanilsaha.com/blog/filtering-after-top-k-silent-rag-bug/ https://swapnanilsaha.com/blog/filtering-after-top-k-silent-rag-bug/ Mon, 21 Sep 2026 09:00:00 +0530 Fetch the nearest k, then filter: a filtered vector search comes back short and nothing says so. The math, four real instances, and the fixes that hold. https://swapnanilsaha.com/blog/should-a-correction-repeat-the-mistake/ https://swapnanilsaha.com/blog/should-a-correction-repeat-the-mistake/ Fri, 18 Sep 2026 09:00:00 +0530 Agent memory can delete a wrong belief, replace it, or keep it labelled false. Human research says restate the myth; transformer research says that may prime it. https://swapnanilsaha.com/blog/agents-forget-how-not-what/ https://swapnanilsaha.com/blog/agents-forget-how-not-what/ Tue, 15 Sep 2026 09:00:00 +0530 Agent memory stores facts about code. In real bugfix runs the missing piece was how to get an honest test result, and when that knowledge is absent nothing fails. https://swapnanilsaha.com/blog/forgetting-is-a-capability/ https://swapnanilsaha.com/blog/forgetting-is-a-capability/ Fri, 11 Sep 2026 09:00:00 +0530 Every memory system optimizes retention. Generalization is lossy compression, so a system that keeps every particular cannot abstract. The operation that fixes this is missing. https://swapnanilsaha.com/blog/identity-is-made-of-memory/ https://swapnanilsaha.com/blog/identity-is-made-of-memory/ Wed, 09 Sep 2026 09:00:00 +0530 An agent has no way to be the same agent twice. Without a "self", nothing decides what is worth remembering. The two problems are one. https://swapnanilsaha.com/blog/consolidation-is-the-missing-phase/ https://swapnanilsaha.com/blog/consolidation-is-the-missing-phase/ Mon, 07 Sep 2026 09:00:00 +0530 Every answer to 'not enough context' is more context. Biological memory never scaled capacity. It built an offline phase that decides what survives. https://swapnanilsaha.com/blog/the-writer-cannot-know-the-readers-question/ https://swapnanilsaha.com/blog/the-writer-cannot-know-the-readers-question/ Sat, 05 Sep 2026 09:00:00 +0530 Every note is written before the query that will need it exists. That makes write-time precision undefined, and the fix is the unit of storage, not a better guess. https://swapnanilsaha.com/blog/recall-is-a-write-operation/ https://swapnanilsaha.com/blog/recall-is-a-write-operation/ Tue, 01 Sep 2026 09:00:00 +0530 Retrieving a memory changes what gets retrieved next. That feedback loop makes early accidents permanent and hides the evidence that would correct them. https://swapnanilsaha.com/blog/a-guess-becomes-a-fact/ https://swapnanilsaha.com/blog/a-guess-becomes-a-fact/ Mon, 31 Aug 2026 09:00:00 +0530 A user instruction, a tool result, a file fact, and the model's own guess end up as the same type once they leave the context window. Forgetting costs a re-derivation. Laundering corrupts everything downstream. https://swapnanilsaha.com/blog/relevance-is-the-wrong-sort-order/ https://swapnanilsaha.com/blog/relevance-is-the-wrong-sort-order/ Sun, 30 Aug 2026 09:00:00 +0530 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. https://swapnanilsaha.com/blog/absence-is-the-memory-nobody-writes/ https://swapnanilsaha.com/blog/absence-is-the-memory-nobody-writes/ Sat, 29 Aug 2026 09:00:00 +0530 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. https://swapnanilsaha.com/blog/a-directive-is-not-a-memory/ https://swapnanilsaha.com/blog/a-directive-is-not-a-memory/ Fri, 28 Aug 2026 09:00:00 +0530 Obeying a rule destroys the evidence that would test it. Agent memory stores directives next to observations and applies the same checks, which all miss. https://swapnanilsaha.com/blog/reinstatement-is-the-missing-primitive/ https://swapnanilsaha.com/blog/reinstatement-is-the-missing-primitive/ Wed, 26 Aug 2026 09:00:00 +0530 Every agent memory system can delete or supersede a belief. Almost none can un-revoke one, and that missing edge manufactures self-reinforcing error. https://swapnanilsaha.com/blog/search-is-memorys-write-path/ https://swapnanilsaha.com/blog/search-is-memorys-write-path/ Tue, 25 Aug 2026 09:00:00 +0530 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. https://swapnanilsaha.com/blog/agents-over-trust-instructions-and-under-trust-corrections/ https://swapnanilsaha.com/blog/agents-over-trust-instructions-and-under-trust-corrections/ Mon, 24 Aug 2026 09:00:00 +0530 Agents obey planted instructions in memory files and ignore true corrections in the same files. Trust in ambient context is not a scalar, it is shaped. https://swapnanilsaha.com/blog/verifying-ai-subagent-work/ https://swapnanilsaha.com/blog/verifying-ai-subagent-work/ Wed, 19 Aug 2026 09:00:00 +0530 A subagent's completion report is testimony, not evidence. Nine mechanical, adversarial checks that turn the claim into a fact, with real cases. https://swapnanilsaha.com/blog/agent-memory-legibility-problem/ https://swapnanilsaha.com/blog/agent-memory-legibility-problem/ Tue, 18 Aug 2026 09:00:00 +0530 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. https://swapnanilsaha.com/blog/agent-memory-expiry-is-a-state-not-a-delete/ https://swapnanilsaha.com/blog/agent-memory-expiry-is-a-state-not-a-delete/ Sun, 16 Aug 2026 09:00:00 +0530 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. https://swapnanilsaha.com/blog/agent-eval-leaks/ https://swapnanilsaha.com/blog/agent-eval-leaks/ Sat, 01 Aug 2026 09:00:00 +0530 A control arm that should have known nothing named the hidden mechanism out loud. It read it off its own working directory. Five leak classes and the fixes. https://swapnanilsaha.com/blog/dogfooding-bugs-ai-memory-daemon/ https://swapnanilsaha.com/blog/dogfooding-bugs-ai-memory-daemon/ Thu, 30 Jul 2026 09:00:00 +0530 Six real bugs no test suite and no reviewer caught. Only residency did. Each one is a boundary a test harness has no way to represent. https://swapnanilsaha.com/blog/agent-memory-trust-ladder/ https://swapnanilsaha.com/blog/agent-memory-trust-ladder/ Tue, 28 Jul 2026 09:00:00 +0530 Three ways a stored note lies, the trust ladder I shipped to handle them, and the three problems I have not solved. https://swapnanilsaha.com/blog/sentinel-pattern-multi-agent-development/ https://swapnanilsaha.com/blog/sentinel-pattern-multi-agent-development/ Sun, 26 Jul 2026 09:00:00 +0530 One orchestrator that writes no product code, disposable coders in isolated worktrees, reviewers paid to attack, and seven rules that each cost me something to learn. https://swapnanilsaha.com/blog/agent-memory-harness-property/ https://swapnanilsaha.com/blog/agent-memory-harness-property/ Fri, 24 Jul 2026 09:00:00 +0530 Every shipped agent memory design assumes the model volunteers to save and recall. In a seeded control it volunteered zero times in 114 turns. What to build instead. https://swapnanilsaha.com/blog/mcp-tool-adoption-agents/ https://swapnanilsaha.com/blog/mcp-tool-adoption-agents/ Mon, 20 Jul 2026 09:00:00 +0530 A controlled study of MCP tool adoption: 0 calls in 63 turns with 30 mentions of the tool — and what actually moves the number. https://swapnanilsaha.com/blog/deleting-ranking-heuristics-code-search/ https://swapnanilsaha.com/blog/deleting-ranking-heuristics-code-search/ Sun, 19 Jul 2026 09:00:00 +0530 3,817 lines of if/else, deleted in one commit — and the measured evidence for why query-keyword branches can never rank code search. https://swapnanilsaha.com/blog/embedding-dilution-code-search/ https://swapnanilsaha.com/blog/embedding-dilution-code-search/ Sat, 18 Jul 2026 09:00:00 +0530 The docstring paraphrased the query and it still ranked below 200 look-alikes. A measured post-mortem of embedding dilution — and the dual-vector fix it shipped. https://swapnanilsaha.com/blog/what-survives-compact-claude-code/ https://swapnanilsaha.com/blog/what-survives-compact-claude-code/ Fri, 17 Jul 2026 09:00:00 +0530 We forced 100+ compactions and graded, fact by fact, what the boundary keeps and what silently dies. The measured map, honest costs included. https://swapnanilsaha.com/blog/vectr-1-1-team-mode-seven-agent-test/ https://swapnanilsaha.com/blog/vectr-1-1-team-mode-seven-agent-test/ Sat, 11 Jul 2026 09:00:00 +0530 Team mode, auth, and encryption at rest — gated by a test where seven AI agents built and reviewed a product coordinating only through vectr's shared memory. https://swapnanilsaha.com/blog/vectr-v1-release-gate-honest-numbers/ https://swapnanilsaha.com/blog/vectr-v1-release-gate-honest-numbers/ Wed, 08 Jul 2026 09:00:00 +0530 A first-person dogfood gate caught two release-blocking bugs before the tag — and the honest cost numbers, wins and losses, that shipped with v1.0.0. https://swapnanilsaha.com/blog/four-families-llm-context-relief-eviction/ https://swapnanilsaha.com/blog/four-families-llm-context-relief-eviction/ Mon, 06 Jul 2026 09:00:00 +0530 Eviction, offload-and-recall, retrieval-over-stuffing, and subagent isolation — how to keep a coding agent from drowning in its own context, and why the four only work when composed. https://swapnanilsaha.com/blog/claude-code-hooks-deterministic-agent-memory/ https://swapnanilsaha.com/blog/claude-code-hooks-deterministic-agent-memory/ Sat, 04 Jul 2026 09:00:00 +0530 Events, the settings.json contract, exit codes — and a real production hook pipeline that injects working memory into an agent deterministically. https://swapnanilsaha.com/blog/building-vectr-part-3-benchmark-methodology-results/ https://swapnanilsaha.com/blog/building-vectr-part-3-benchmark-methodology-results/ Sun, 14 Jun 2026 09:00:00 +0530 How I benchmarked an AI code editor tool without fooling myself — methodology, the research vs implementation distinction, and what the CPython results actually show. https://swapnanilsaha.com/blog/building-vectr-part-2-working-memory-compact-survival/ https://swapnanilsaha.com/blog/building-vectr-part-2-working-memory-compact-survival/ Thu, 11 Jun 2026 09:00:00 +0530 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 made it real. https://swapnanilsaha.com/blog/building-vectr-part-1-semantic-code-search/ https://swapnanilsaha.com/blog/building-vectr-part-1-semantic-code-search/ Tue, 09 Jun 2026 09:00:00 +0530 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. https://swapnanilsaha.com/blog/llm-context-window-token-budget-guide/ https://swapnanilsaha.com/blog/llm-context-window-token-budget-guide/ Tue, 26 May 2026 09:00:00 +0530 Token budget arithmetic for LLM context windows: why system prompts, tool schemas, RAG, and history drain your window — and four strategies to manage it. https://swapnanilsaha.com/blog/why-ai-code-assistants-waste-context-rag/ https://swapnanilsaha.com/blog/why-ai-code-assistants-waste-context-rag/ Sun, 24 May 2026 09:00:00 +0530 Context window stuffing silently degrades AI code assistant accuracy. Learn exactly why attention dilutes over long inputs and how RAG for code fixes it. https://swapnanilsaha.com/blog/text-embeddings-llms-rag-complete-guide/ https://swapnanilsaha.com/blog/text-embeddings-llms-rag-complete-guide/ Sat, 23 May 2026 09:00:00 +0530 From tokenization and transformers to cosine similarity, HNSW, RAG pipelines, and LLM training costs. With 7 interactive demos. https://swapnanilsaha.com/blog/dpdp-act-ai-llm-automation/ https://swapnanilsaha.com/blog/dpdp-act-ai-llm-automation/ Fri, 22 May 2026 09:00:00 +0530 What the DPDP Act actually requires, why shared inboxes fail, and how AI classifies requests, drafts multilingual replies, and generates audit-ready evidence — a deep technical guide.