MCPMemory for Claude Code, Cursor & Codex

Stop re-explaining your codebase to AI.

Your coding agent forgets everything between sessions. Recira gives it a persistent memory of your repo's decisions, conventions and lineage — extracted automatically, served over MCP, recalled in milliseconds.

p95 142ms recall< 2k tokens / answerzero-leak tenant isolation
agent session — recira mcplive
agent
Ingests from
The problem

Every session starts from zero.

Context windows are short-term memory. The moment a session ends, your agent forgets the architecture it just learned, the decision it just made, and the convention you just corrected.

You
mondayUse ULIDs for new primary keys, not UUIDs — we standardized on that.
AI
agentGot it — switching to ULIDs.
AI
thursday · new sessionI'll use uuid4() for the primary key here…

It re-derives what you already decided

Architectural decisions, trade-offs and dead-ends live in PRs and postmortems the agent never sees. So it proposes the approach you rejected six months ago.

You pay the context tax every time

Re-pasting files, re-explaining conventions, re-stating the plan. Tokens, latency and your attention — burned re-teaching the same things on repeat.

Knowledge walks out the door

When the engineer who made the call leaves, the why goes with them. Your agents inherit the code, but none of the reasoning behind it.

How it works

Ingest. Extract. Recall.

Recira reads your engineering knowledge, distills it into atomic memory, and hands it back to your agent exactly when it's writing code.

1

Ingest your sources

Connect a repo and Recira pulls the full history — code, commits, PR threads, ADRs — plus Confluence, Slack, Google Docs and Notion. Public repos start indexing instantly.

gitconfluenceslack+3
2

Extract atomic memory

A cost-tiered cascade distills raw docs into MemCells — single propositions like a decision or convention — each with a source, confidence and validity over time. 71% never touch an LLM.

tree-sitterbi-temporalconfidence
3

Recall over MCP

Your agent calls Recira like any other tool. It gets the few decisions and conventions that matter for the task — scoped, ACL-filtered, and under 2k tokens — in ~142ms.

mcpscopedp95 142ms
Sources
6 connectors
Extract
cost cascade
MemCells
8.4M indexed
Your agent
over MCP
Integrations

Drops into the editor you already use.

Recira is a standard MCP server. Paste one config block and your agent gains a memory — no plugins, no proxies, no rewiring your workflow.

recall_decisions_about
Decision lineage for a symbol, file or topic.
recall_failed_attempts
Approaches that were tried and reverted — so the agent doesn't repeat them.
surface_conventions
Active conventions in scope before it writes a line.
remember_fact
Write a fact mid-session — memory that compounds as you work.
~/.cursor/mcp.json
{
  "mcpServers": {
    "recira": {
      "url": "https://api.recira.ai/mcp/mcp",
      "headers": {
        "Authorization": "Bearer rc_live_8f3x…a91d"
      }
    }
  }
}

Long-lived, revocable token. Scoped to your workspace. Every retrieval respects source-system ACLs.

Why Recira

One decision. Every source.

Other tools index files. Recira reconstructs reasoning — linking a single decision across the RFC, the Slack thread, the PR and the postmortem into one traceable lineage.

CONFLUENCE · RFC
“Evaluate SCIP vs tree-sitter for symbol extraction.”
2026-02-04 · design doc
SLACK · #eng-memory
“tree-sitter is fast enough for the hot path.”
2026-02-06 · thread
GIT · PR #412 — merged
Decision: tree-sitter over SCIP — 94% quality at ~0 cost.
2026-02-09 · conf 0.86
GDOCS · POSTMORTEM
Cross-file refs regression → SCIP retained for refs only.
2026-03-01 · follow-up

Cross-source lineage

The full causal chain behind a decision — not just the line of code that resulted from it.

Bi-temporal truth

Recira knows what was true when. Superseded decisions are kept and marked — never silently overwritten.

Zero-leak isolation

Tenant binding comes from the token, never a parameter. Every retrieval is filtered by your real source-system ACLs.

Engineered for cost

A strict envelope — ≤ $0.10 per 100k-word write, ≤ $0.001 per query — holds at any scale.

142ms
p95 RECALL LATENCY
71%
WRITES NEVER HIT AN LLM
<2k
TOKENS PER ANSWER
8.4M
MEMCELLS SERVED
Get started

Give your agents a memory.

Connect a repo in minutes and watch your coding agent stop asking the same questions twice. We're onboarding teams now.

hello@recira.ai · SOC 2 Type II · works with Claude Code, Cursor & Codex