Unified Agent Context

English | 한국어 | 中文 | 日本語 | Español

Every new AI chat starts with amnesia. Yesterday you told Claude you like your READMEs in English. This morning Codex asked. Tonight a third assistant will ask again. You run five AI agents, and you are the only memory they share. You are spending yourself one re-explanation at a time.

Unified Agent Context ends that. Tell any agent a decision once. Every other agent knows it in its very next session. Automatically, on every machine, with secrets locked out by design.

You never wire memory together by hand. You never copy context from one agent to another. That is the whole promise. If you ever have to do it yourself, UAC has failed.

How it works

Agents ──(MCP stdio, over ssh when remote)──► mcp-memory-keeper (canonical: sov ~/.uac/data/memory)
   │  ▲
   │  └─ Session start: inject distilled facts only (hook or instructed pull) — scripts/inject-context.mjs
   └──── During session: explicit record (record-fact) / on exit: auto-distill (distill-session)
              └─ Every write path goes through a central fail-closed secret gate (block or redact)
              └─ Permanent facts queue into the librarian outbox → librarian-sync delivers to the
                 sov Letta "The Noticer" inbox (Phase 5)

In plain words.

• All your agents read and write one shared memory.

• A preference follows you everywhere. A project decision stays inside its project.

• Decisions and preferences are kept forever. Day to day work state expires after 90 days.

• Raw conversations are never injected into sessions. They rest in a cold archive, and secrets are scrubbed on the way in.

• Every write passes a secret gate. Keys and passwords are blocked by default.

• If the store is unreachable, agents say so loudly instead of guessing quietly.

• The main store lives on one home server named sov. Every other machine reaches it over ssh.

Related MCP server: persistent-kb-mcp

The docs

The commands

node scripts/inject-context.mjs [--cwd <dir>]        # print the shared context block
node scripts/record-fact.mjs --type decision "..."   # explicit record (secrets blocked with exit 3)
node scripts/distill-session.mjs --file <transcript> # distill a session + quarantine sweep
node scripts/librarian-sync.mjs [--strict]           # deliver outbox → sov librarian inbox (Phase 5)
node scripts/doctor.mjs                              # wiring self-check
node scripts/cross-verify.mjs                        # re-run the 20-path coverage matrix
node scripts/reexplain.mjs log|report                # re-explanation metrics (2-week gate aux indicator)
node --test 'tests/*.test.mjs'                       # full test suite (73)

The two week test

All the scenario tests pass. The real test is different. After two weeks of daily use, do you still catch yourself explaining the same thing twice. Log every repeat with reexplain.mjs log. If the weekly count falls to zero, UAC works.

Credits

Built end to end with GJC, Gajae Code, an AI coding agent. The store phases, the story style of this README in five languages, and the promo animation above were implemented, verified, and shipped by GJC. While building UAC it was using the shared memory of UAC itself.