Vector: Local Working Memory for Coding Agents

Compact, private, zero-model-call recall for agents that need continuity without re-reading everything.

Comment

One-line pitch

Vector gives coding agents compact, local working memory and sub-second recall without spending model tokens re-reading an entire project.

Description

Vector is a local, deterministic memory and recall engine built with Codex on an older Intel MacBook. It turns project material into structured atoms and returns small ranked working-memory packs before an agent acts. During Build Week, Codex added non-cold AIRLOCK continuity, bounded scanner recovery, an incremental dense ANN, proactive capsule indexing, hybrid recall, and a read path that avoids neural-model startup. Apple semantic embeddings stay on the write path; normal recall remains local and requires no model or network call.

Codex collaboration

Codex/GPT-5.6 mapped the existing engine, implemented the bounded Build Week additions, wrote verifiers and benchmarks, diagnosed scanner stalls, and made failures visible. Marty directed the architecture: preserve deterministic local recall, bound read-time cost, separate private data from judge material, and do not hide failed gates.

Judge path

python3 judge_demo.py

The sanitized harness uses synthetic atoms and requires no API key, model, account, network connection, private corpus, or native rebuild.

Current canonical evidence

  • Canonical HEAD: 765810d
  • Prior-work baseline: 7b8d1329f10ad3227ec0ed82d244ce66c5236fb7
  • Build Week commits: 51
  • Canonical verifier: PASS 21/21
  • Fresh-clone test: PASS
  • Live corpus scale: 31,386 vectors
  • Live index scale: 125,544 buckets
  • Internal recall: 241.4 ms
  • Process wall time: 446.1 ms
  • Recall-time model/API calls: 0
  • Query-time model tokens: 0

The zero-call and zero-query-token claims apply only to local Vector recall, not to downstream model generation.

Demo

The public 2:55 demo, “Build week Vector Demo,” shows local query recall, compact ranked context, operation on a 10-year-old Intel MacBook, token savings at recall time, and how Codex built the Build Week additions.

Prior work and scope

The prior-work baseline separates work completed before Build Week. The Build Week sequence contains AIRLOCK continuity, scanner hardening, dense ANN, proactive indexing, token hierarchy wiring, and no-neural-read changes.

The private corpus and spatial integration source are intentionally excluded from the judge package.

Sessions

Primary feedback session:

019f7279-6ec0-7c62-88b3-414259b27ce3

Support session:

019f82f5-316d-7693-99c5-9cf9ea731b54

Track

Developer Tools

Built With

  • apple-semantic-embeddings
  • codex
  • gpt-5.6
  • python
Share this project:

Updates