Vectr · Agent Memory Research · Deep-Dive Writing

Swapnanil Saha

I build Vectr, semantic search and working memory for AI coding agents, and write long-form engineering deep dives. The research underneath it is on arXiv.

Read the writing Vectr The paper
Latest Deep Dive — 20 July 2026
Your MCP Tool Works. The Model Still Won't Call It.
0 calls in 63 turns despite 30 mentions of the tool — and the harness injection that flipped it · 19 min read →
Deep-dive essays
15
Retrieval · Memory · Agents
Vectr
v1.5
Search + memory for agents · PyPI →
Compactions instrumented
246
Delivery, Not Storage · the paper →
Tool calls saved
−72%
Vectr on the Apache Camel benchmark →
Writing

Deep dives.
Measured, not summarized.

Long-form engineering essays on retrieval, agent memory, and context engineering — built on instrumented runs and honest ledgers, with every number traceable to an artifact.

19 July 2026 · 20 min · Code Search

I Deleted Every Ranking Heuristic From My Code Search Engine

3,817 lines of query-side if/else, deleted in one commit — and the measured evidence for why keyword branches can never rank code search.
18 July 2026 · 21 min · Embeddings

Embedding Dilution: Why Semantic Code Search Misses the Answer

The target's docstring paraphrased the query almost word for word, yet ranked below 200 look-alikes — a measured post-mortem and the dual-vector fix.
17 July 2026 · 28 min · Claude Code

What Actually Survives /compact in Claude Code

108 + 138 forced compactions, graded token by token. What the boundary keeps, what silently dies, and where to put state so it survives.
All 15 essays RSS swapnanilsaha.com/blog →
Vectr

Semantic search.
Working memory.

Open source · MCP · Zero config
Vectr gives AI coding agents the two things they lack: retrieval that understands code, and memory that survives.
An AST-aware codebase indexer with hybrid semantic + BM25 search, a symbol graph, and a persistent working-memory tier delivered through hooks — findings survive context compaction and session restarts instead of dying with the conversation. Any MCP-compatible AI code editor connects automatically. v1.1 added team mode — many agents, one shared memory; v1.3 added per-memory triggers, so a note resurfaces exactly when its path, symbol, or lifecycle condition matches; v1.5 adds automatic episode capture, where the harness records what was tried and what failed, and distils the recurring arcs into notes.
Vectr in detail GitHub The research
llm-tools suite

Six tools.
Shipped.

Six production-grade tools — each solving a real enterprise problem with LLM APIs. Built independently, in public, to the same standard I'd apply to internal tooling at scale. Each ships with Docker, a REST API, input validation, and a full landing page.

Developer Tools · Code AI · MCP

vectr

Zero-config codebase indexer for AI assistants. AST-aware chunking, hybrid semantic + BM25 search, and cross-session memory via MCP. Your AI code editor — Claude Code, Cursor, or any MCP client — gets only the relevant chunks, not entire files. Benchmarked on Apache Camel: −58% Phase 2 cost, −72% tool calls.
Compliance · DPDP Act · India

dpdp-copilot

Self-hosted operator tool for India's DPDP Act. Public intake form accepts data principal requests in 5 Indian languages. Automatically classifies and drafts a compliant multilingual reply. Operator inbox shows live SLA status. Every action logged as immutable evidence — exportable as PDF or CSV for regulatory audit.
AI Architecture · RAG

rag-readiness

Opinionated architecture recommendation per component. Diagnoses existing broken stacks with root causes by severity. Rule-based cost estimation, iterative refinement with session persistence, and RAGAS eval dataset generation.
LLM · Evaluation · Regression · CI/CD

llm-eval-suite

Evidence-backed multi-dimensional scoring across 10 task types including RAGAS-compatible RAG eval. Regression testing saves baselines and blocks CI when scores drop. Claim-level hallucination detection, prompt sensitivity analysis across variants, and multi-judge panel consensus with disagreement flagging.
Enterprise AI · Strategy · LangGraph

ai-use-case-scoper

Hybrid 8-question flow with optional Graph RAG doc enrichment — pre-fills tech context, always asks qualitative questions. LangGraph check-in pipeline evolves plans. Company memory persists across sessions.
Productivity · Meeting AI

meeting-to-action

Normalizes Zoom/Teams/Meet transcripts, detects meeting type, extracts structured decisions and action items. Cross-meeting commitment tracker flags missed items by severity. Weekly digest aggregates open work by owner across all sessions.
All tools open-source · github.com/swapnanil
Now

What I'm
building.

Active · Vectr
Shipping the memory half of Vectr
v1.5 shipped automatic episode capture with arc distillation, and a note lifecycle where a wrong note is revoked rather than deleted, so the mistake is not silently re-derived, on top of 1.3's per-memory trigger engine and 1.1's team mode. The research underneath it is now published: Delivery, Not Storage (arXiv:2607.20972), a controlled study of agent memory under repeated context compaction across 108 and 138 forced boundaries, graded token by token. Running alongside it: an upstream feature request to Claude Code (#78795) proposing triggered injection for auto-memory files, so the findings become platform behavior rather than a workaround.
Writing · Next up
Next deep dives in the pipeline
The arXiv paper rewritten as a readable deep dive, and the cue-anchored trigger vocabulary taken apart condition by condition. Each one built on measured runs, like everything else here.
Retrieval research
Understanding-grade code retrieval
Moving Vectr's search from embeddings-plus-heuristics to structural signals: per-symbol purpose vectors shipped in v1.0, a file-level reference-importance prior ships in the ranking layer, and the class-level owning-class prior is next, each one evidence-gated on public corpora.
About

Where the instinct
comes from.

I build Vectr, write long-form engineering deep dives, and publish the research underneath them. Beneath all of it sit nine years in the engine room of real-time digital advertising, building infrastructure that decides in milliseconds which ad a user sees: a custom Redis module that cut filtering latency 5×, first-price and second-price auction engines, real-time budget pacing on Cassandra across billions of daily requests.

Ad-tech has an unusually short feedback loop. Every latency improvement carries a win-rate implication, every auction change shows up in revenue by the same afternoon. That is where the habit of refusing a claim without an instrumented run behind it comes from, and it is the standard everything published here is held to.

Building
Vectr · The deep-dive blog
Research
Agent memory · Context engineering
Background
Real-time systems · Ad-tech
Location
Mumbai, India
Contact

Let's
talk.

Open to conversations about agentic AI, context engineering, retrieval, and the harness problems that surface when a model has to hold state across long-running work. Early results, half-formed hypotheses, and work that has not been written up yet are the most interesting kind of mail to get, and the topic does not have to be one covered here.

Corrections carry the same weight as agreement. A number on this site that does not reproduce is worth an email.

Email swapnanilsaha26@gmail.com GitHub github.com/swapnanil LinkedIn linkedin.com/in/swapnanil
Most of the interesting problems in this field are still unsolved and largely undocumented. Writing them down carefully, with the run attached, is how they stop being folklore.