Queryable memory for code

Give your agents total recall.

code.subspace turns the history scattered across repositories, branches, and worktrees into durable context your coding agents can query. Less searching. Fewer repeated mistakes. Better answers backed by the code that came before.

Explore the queries

cross-repository history-aware SQL + Datalog

agent session · code.subspace

Payments started double-charging Tuesday. Find the change across every repository.

SELECT repo, unit, author, arrived
FROM search_history('payment retry idempotency')
WHERE arrived >= '2026-03-01'
ORDER BY relevance DESC;

repo	unit	author	arrived
payments-svc	retry_charge	m.tanaka	2026-03-01
web-app	useRetryPayment	d.kim	2026-03-02

retry_charge creates a new idempotency key on every retry. That is the regression.

ONE QUERY SURFACEMeaning, structure, provenance, and time in the same answer.

HISTORY THAT STAYS USEFULRecall work after branches are deleted or commits are rewritten.

CONTEXT ACROSS THE FLEETSee relevant code across repositories and active worktrees.

Why it works

Code history becomes working memory.

Agents are excellent at reasoning over context. The hard part is finding the right context without flooding the window. code.subspace reduces a sprawling code estate to focused, composable results.

⌕

Ask by meaning, narrow by evidence

Start with semantic similarity, then join against authorship, callsites, branches, repositories, or arrival time. The result is precise enough to act on and small enough to reason about.

Identity follows the code

Content-addressed units connect equivalent code across paths and repositories, making moves, copies, and shared implementations visible without brittle path matching.

↺

Dead ends remain instructive

Abandoned experiments and deleted branches can still answer “have we tried this?” Agents learn from discarded approaches instead of quietly repeating them.

Δt

Timelines reflect when code arrived

Append-only ingestion time separates the history your team received from the dates encoded in commits. That makes incident investigation resilient to rebases and force pushes.

Query gallery

One index. Many kinds of answers.

Purpose-built views keep each query focused on the agent’s intent. Switch between SQL and Datalog to see the same task expressed in either language.

“Sessions started expiring after yesterday’s deploy. What new code touched token or TTL logic in any repository?”

SELECT repo, unit, author, arrived
FROM search_history('session token ttl expiry')
WHERE arrived > now() - INTERVAL '72 hours'
ORDER BY relevance DESC;

[:find ?repo ?name ?author ?arrived
 :where
 (search-history "session token ttl expiry"
                 ?repo ?name ?author ?arrived)
 (within-last ?arrived "72h")]

repo	unit	author	arrived
auth-svc	clamp_token_ttl	p.iyer	26 h
shared-lib	CACHE_TTL_S	d.kim	41 h

Root cause without repository-by-repository archaeology. The time filter reflects arrival in the index, so rewritten commit dates do not distort the investigation window.

“Did anyone try building a feature-flag SDK? Show abandoned approaches, newest first.”

SELECT unit, author, branch, last_seen
FROM search_history('feature flag sdk')
WHERE alive_at_head = false
ORDER BY last_seen DESC;

[:find ?name ?author ?branch ?last-seen
 :where
 (search-history "feature flag sdk" ?unit)
 (not (alive-at-head ?unit))
 (last-ref ?unit ?branch)
 (unit-summary ?unit ?name ?author ?last-seen)]

unit	author	last branch	last seen
FlagClient.evaluate	s.novak	spike/flags-v2	2026-02-19
FeatureGate	d.kim	feat/gates	2024-09-30

Discarded work becomes institutional memory. The agent can compare old approaches before designing another version from scratch.

“Is another agent already implementing rate limiting in an unmerged worktree?”

SELECT agent, branch, unit, pushed_ago
FROM search_worktrees('token bucket rate limiter')
ORDER BY relevance DESC
LIMIT 3;

[:find ?agent ?branch ?name ?age
 :where
 (search-worktrees "token bucket rate limiter" ?unit)
 (unmerged ?unit ?agent ?branch)
 (unit-summary ?unit ?name ?age)]

agent	branch	unit	pushed
agent-7	wt/rl-middleware	RateLimiter.allow	14 min
agent-7	wt/rl-middleware	TokenBucket.take	14 min

Coordination happens before the collision. Agents can discover overlapping work early, build on it, or choose a different task.

“Trace validate_webhook through every rename, move, and repository where it lived.”

SELECT repo, path, unit, first_seen
FROM unit_lineage('validate_webhook')
ORDER BY first_seen;

[:find ?repo ?path ?name ?first-seen
 :where
 (lineage "validate_webhook" ?unit)
 (unit-location ?unit ?repo ?path ?name)
 (first-seen ?unit ?first-seen)]

repo	path	unit	from
api-gateway	src/hooks.py	verify_signature	2023-04-11
shared-lib	pkg/webhook/verify.py	validate_webhook	2025-08-21

Blame without path boundaries. Content identity ties a unit to its earlier forms even after it moves into a shared library.

Architecture

Connect once. Ask from anywhere.

Local and remote hooks feed a shared, content-addressed model of your code. Agent tools turn natural-language intent into focused SQL or Datalog.

Ingest changes

Capture code units, commit context, and arrival time across connected repositories and worktrees.

Connect the history

Unify semantic embeddings, unit identity, provenance, and repository structure in one model.

Return evidence

Agents receive compact rows and precise locations instead of another full-repository dump.

service-apimain + feature branches

web-clientcommits + active worktrees

shared-libshistory + provenance

code.subspacesemantic index
content identity
commit graph
arrival timeline

agent toolsMCP · SQL · Datalog

Deployment

Run it where your code belongs.

Choose infrastructure you operate or a managed service. Both paths expose the same query model and agent-facing tools.

MIT-licensed core

Self-hosted

Deploy the engine with Postgres, your preferred embedding model, and your object store. Keep control of infrastructure, retention, and access policy.

Managed cloud

code.subspace Cloud

Connect organization repositories without operating the ingestion and vector fleet. Designed for shared agent access, SSO, and centralized retention controls.

Stay in the loop

Let your next agent start with context.

Join the newsletter for product updates, technical notes, and invitations to the early-access program.