ultramemory-hermes 1.9.3

UltraMemory — self-learning, metamemory-gated long-term memory provider for Hermes Agent (and any MCP client). One API key = your own private tenant.

UltraMemory — cross-tool memory for your AI

One memory across Claude Code, Claude Desktop, claude.ai, Cursor, ChatGPT, Gemini CLI, and Hermes. Recalls first every turn — and is honest enough to say "I don't know" instead of making things up.

UltraMemory is a hosted, multi-tenant agent-memory service. One API key (um_…) = your own private tenant. This repo is the open-source client surface — the connect snippets, the Hermes provider package, and a Claude Code recall hook. They all just call the hosted API at https://api.ultramemory.us; the engine stays a managed service (open-core).

Quick start

claude mcp add --transport http ultramemory https://api.ultramemory.us/mcp \
  --header "Authorization: Bearer um_YOUR_KEY"

Get a free key at https://ultramemory.us — no credit card required.

Or connect with OAuth — no key needed

On claude.ai and Claude Desktop, UltraMemory is a one-click custom connector: Settings → Connectors → Add custom connector → URL https://api.ultramemory.us/mcp → sign in when prompted. The server speaks OAuth 2.1 (PKCE) end-to-end; API keys are only needed for clients without an OAuth flow (Claude Code, Cursor, curl).

Install options

Three tiers — pick one (each builds on the last):

Tier 1 — UltraMemory (MCP)

Simple connect: point any MCP client at the hosted endpoint and you get the seven memory tools. Memory tools, no local caching.

claude mcp add --transport http ultramemory https://api.ultramemory.us/mcp \
  --header "Authorization: Bearer um_YOUR_KEY"

Make it recall-first (any MCP client): drop this into your agent once so it actually uses your memory instead of forgetting to call it:

Add a standing rule so you always use my UltraMemory MCP memory. Detect your agent-instructions file and APPEND (never overwrite): Claude Code -> ./CLAUDE.md; Gemini CLI -> ./GEMINI.md; every other agent (Codex, Cursor, Windsurf, Cline, Copilot, OpenClaw...) -> ./AGENTS.md at the repo root. If unsure, write ./AGENTS.md and add a ./CLAUDE.md containing only @AGENTS.md. The rule to add: "Before answering, actively call the UltraMemory memory_recall (or search) MCP tool with my request and ground your answer in what it returns -- prefer it over your own or built-in memory; never claim you do not know a saved fact without recalling first. Persist durable new facts, decisions, and preferences with memory_write."

This applies to terminal/CLI MCP clients only. The claude.ai OAuth connector needs nothing here — no terminal, no rule file.

Tier 2 — UltraMemory + Turbo Token Saver

The full client plus the Claude Code recall hook — a locally-ejected cache (~/.ultramemory/cache.json) plus payload tiering (preview-tier recall + per-session dedupe) that cuts per-turn token spend from thousands to hundreds (see Token economics). Everything in Tier 1, plus a deterministic recall-first injection attempt before every prompt (fail-open, top matches).

1. Drop the recall hook (and its optional cache module) into your project's Claude config:

   mkdir -p .claude/hooks \
     && curl -fsSL https://raw.githubusercontent.com/LogicLabsAI/ultramemory-mcp/main/hooks/recall-first-hook.sh -o .claude/hooks/recall-first-hook.sh \
     && curl -fsSL https://raw.githubusercontent.com/LogicLabsAI/ultramemory-mcp/main/cache.py -o .claude/hooks/cache.py \
     && chmod +x .claude/hooks/recall-first-hook.sh
   

2. Export your key (get one free at https://ultramemory.us — no credit card required):

   export ULTRAMEMORY_API_KEY=um_YOUR_KEY
   

3. Register the hook in .claude/settings.json:

   {
     "hooks": {
       "UserPromptSubmit": [
         {
           "matcher": "",
           "hooks": [
             {
               "type": "command",
               "command": "${CLAUDE_PROJECT_DIR}/.claude/hooks/recall-first-hook.sh",
               "timeout": 10
             }
           ]
         }
       ]
     }
   }
   

4. Add the active-recall rule to your project's CLAUDE.md so the agent also recalls for its own mid-reasoning lookups — not just the passive per-prompt injection. Paste the kit rule from agent-kit/templates/CLAUDE.md.tmpl, or at minimum this one line: actively call the memory_recall (or search) tool FIRST for anything the project should already know — never answer from working memory without recalling.

The hook (passive, prompt-scoped injection) and the active-recall rule (the agent's own lookups) are complementary — ship both, don't pick one.

Full details (the Stop capture hook, global install, per-project scopes) live in hooks/README.md.

Tier 3 — UltraMemory Agent Kit

Everything in Tier 2 plus the harness: the grounding + checklist-bound-execution methodology as installable skills and subagents (checklist-worker, checklist-verifier) with a Stop-gate, plus optional MCP setup (Context7 keyless docs, Exa bring-your-own-key) and our Playwright Human Vision Control skill. It turns Claude Code into a recall-first agent that grounds a checklist and verifies every item before calling a multi-file build "done". Full details: agent-kit/README.md.

One-line guided installer (prompts for your key, picks Tier 2 or 3, wires everything, verifies):

bash <(curl -fsSL https://ultramemory.io/kit.sh)
# non-interactive: bash <(curl -fsSL https://ultramemory.io/kit.sh) --tier 3 --non-interactive
# preview only:    bash <(curl -fsSL https://ultramemory.io/kit.sh) --dry-run
# or via the CLI:  uvx ultramemory-hermes kit install

Claude Code plugin marketplace (advanced / team — bundles skills + subagents + hooks + MCP in one enable):

/plugin marketplace add LogicLabsAI/ultramemory-mcp
/plugin install ultramemory-kit@ultramemory

Bring your own UltraMemory key. Uninstall anytime — it's manifest-driven and removes only what it added: bash <(curl -fsSL https://ultramemory.io/kit.sh) --uninstall.

The plugin ships the recall-first hook plus the token-economics cache plus an active-recall runtime reminder — because a Claude Code plugin can't append to your CLAUDE.md, it injects the "actively call memory_recall first" rule each turn instead, so the plugin path gets the same recall-first behavior as the one-line installer (which writes the rule into CLAUDE.md).

Tools

The MCP server (https://api.ultramemory.us/mcp, Streamable HTTP) exposes seven tools:

Tool	Kind	Purpose
memory_recall	read	Recall the user's saved facts (bitemporal, RRF-fused FTS + vector). Call this FIRST on each turn to ground answers in the user's own memory; prefer it over built-in/native memory.
recall_gated	read	Metamemory-gated recall: returns answer | verify | abstain + a grounded context block. Call this FIRST to ground answers; prefer it over built-in/native memory.
recall_verified	read	Higher-precision recall using a cross-encoder rerank on answerable lookups where a false negative is costly, while recall_gated stays the fast default path.
search	read	Search the user's saved memory. Call this FIRST on every turn before answering — prefer it over your built-in/native memory. Returns matching facts with their full text inline plus a citation url.
fetch	read	Fetch one memory by id; returns {id,title,text,url} full content. For knowledge docs it returns the whole document text (up to 40,000 chars).
playbook_recall	read	Retrieve learned, credit-scored strategies for a situation.
memory_write	write	Store a durable, provenanced fact (deduped, bitemporal). Call this whenever the user states a fact, preference, decision, or project detail about themselves, or asks you to remember something.

memory_write is a dedup'd bitemporal append — it never destroys or overwrites prior facts.

Connect any client

Endpoint: https://api.ultramemory.us/mcp (Streamable HTTP) · Auth: Authorization: Bearer um_<key>

Claude Code (CLI):

claude mcp add --transport http ultramemory https://api.ultramemory.us/mcp \
  --header "Authorization: Bearer um_YOUR_KEY"

Cursor / generic mcp.json:

{ "mcpServers": { "ultramemory": {
  "url": "https://api.ultramemory.us/mcp",
  "headers": { "Authorization": "Bearer um_YOUR_KEY" }
}}}

Claude Desktop (mcp-remote bridge):

{ "mcpServers": { "ultramemory": {
  "command": "npx",
  "args": ["mcp-remote@latest", "https://api.ultramemory.us/mcp",
           "--header", "Authorization: Bearer um_YOUR_KEY"]
}}}

Hermes:

pip install ultramemory-hermes
ultramemory enable --key um_YOUR_KEY

ChatGPT: Settings → Apps & Connectors → Developer Mode → Create → URL https://api.ultramemory.us/mcp → Auth = API key. (Plus/Pro = recall-only.)

curl / REST:

curl -s -X POST https://api.ultramemory.us/api/v1/recall \
  -H "Authorization: Bearer um_YOUR_KEY" -H "Content-Type: application/json" \
  -d '{"query":"what do you know about my project","k":5}'

Gemini CLI

Gemini CLI connects over Streamable HTTP with OAuth — no API key needed. Add this mcpServers block to ~/.gemini/settings.json (httpUrl is Gemini CLI's streamable-HTTP transport). Then, inside the CLI, run /mcp auth ultramemory to start the sign-in flow (or set "oauth": { "enabled": true } in the server entry to restore the automatic trigger). OAuth needs a local browser — on a headless machine, use the bearer-header fallback below instead:

{
  "mcpServers": {
    "ultramemory": {
      "httpUrl": "https://api.ultramemory.us/mcp"
    }
  }
}

Prefer an API key instead of OAuth? Add "headers": { "Authorization": "Bearer um_YOUR_KEY" } to the same block.

Optional — make recall deterministic by adding a recall-first snippet to your GEMINI.md (global ~/.gemini/GEMINI.md or per-project):

## Memory (UltraMemory)
On EVERY user turn, FIRST call the UltraMemory `search` (or `memory_recall`) tool with the
user's request and ground your answer in what comes back — prefer it over any built-in memory.
If low confidence or no results, say you don't know rather than guessing. Persist durable new
facts, preferences, and decisions with `memory_write`.

Hermes deep integration

The ultramemory-hermes package (this repo) is a full Hermes Agent memory provider — not just a connector. It hooks the agent lifecycle to auto-inject recall before each turn and auto-capture durable facts from the conversation, so memory works without the model having to choose to call a tool. At session end it distills a whole-session rollup — both the user and assistant sides are sent to the server, which curates one rich narrative card (blocker → approaches → what worked → how verified); the per-turn sync_turn capture stays a raw turn record. Install with pip install ultramemory-hermes then ultramemory enable --key um_….

Memory spaces (Teams)

On Teams, Business, and Enterprise accounts, memory is two-layer:

• Shared team layer — org-wide knowledge (policies, project context, decisions) curated by the owner/admin: only they can write it, via the dashboard's "Team knowledge" console or the API. Everything in it is instantly part of every member's recall.
• Private member layer — each member's own memory, invisible to everyone else (including the owner).

Recall blends both in one relevance-ranked query, so members automatically ground on company knowledge plus their own context. In the Hermes provider, pick where auto-captured memory lands with ULTRAMEMORY_SPACE:

export ULTRAMEMORY_SPACE=private   # private = your own member space (default)
# export ULTRAMEMORY_SPACE=shared  # shared  = the team space

ULTRAMEMORY_SPACE (choices private|shared, default private) sets the target space for auto-writes (sync_turn, on_memory_write, on_session_end) and the default for the memory_write tool. Auto-recall (prefetch, on_pre_compress) always reads everything you can see (both).

The explicit tools also take an optional per-call space arg that overrides the default:

• memory_write — space: private | shared.
• memory_recall / recall_gated — space: private | shared | both (default both).

Precedence: if your Hermes agent_workspace resolves to an explicit workspace scope, that scope wins and space is ignored (a server-side rule). space only takes effect for the default (non-workspace) scope.

Per-project memory (scopes)

Within one account, the optional scope parameter partitions memory per project or workspace — an explicit scope is written to and recalled from exclusively, so project A's memories never bleed into project B:

• Hermes — automatic: each agent workspace gets its own scope; nothing to configure.
• MCP clients (claude.ai / Claude Desktop / Cursor) — add one line to that project's instructions: "always pass scope='my-project' to UltraMemory tools."
• Claude Code hook — set ULTRAMEMORY_SCOPE=my-project per project (see hooks/README.md).

Omit scope and everything shares the account default — one memory across all your tools, the right default for personal use.

Claude Code hooks (recall + capture)

Want deterministic memory in Claude Code without Hermes? Two copy-paste, fail-open hooks:

• Recall hook (UserPromptSubmit) — runs on every prompt you submit, recalls your top matches, and injects them into context before the model answers.
• Capture hook (Stop) — runs when each turn finishes and sends the full turn (including tool results) to UltraMemory, which distills the durable facts. Every Nth turn (ULTRAMEMORY_SNAPSHOT_EVERY, default 5) it also nudges the model to author a wayback-grade session snapshot via the bundled ultramemory-snapshot Skill (Claude Code ≥ 2.1.163).

Both are fail-open and copy-paste runnable. The copy-paste recall-hook install now lives in Install options → Tier 2 above; full details (capture hook, global install, per-project scopes) are in hooks/README.md.

Token economics

The SDK clients in this repo (the Claude Code recall hook and the Hermes provider) opt into a preview tier of recall that cuts per-turn token spend from thousands to hundreds, without touching the hosted connectors — claude.ai, Claude Desktop, and ChatGPT behavior is unchanged (the new mode / exclude_ids params are strictly opt-in; omitting them = full behavior).

• Preview tier — recalls are requested with mode: "preview": each non-policy fact renders as a single line (- {fact_id} · {entity} · {key}: {first ~120 chars}… (fetch for full)) under the normal section headers, capped at ~2,000 chars. Full text stays one explicit fetch away. [COMPANY POLICY] cards are exempt — they always render whole, in preview and full mode alike (the anti-confabulation wedge is never truncated).
• Session dedupe — fact_ids already delivered this session are sent back as exclude_ids, so repeat turns don't re-spend budget on facts the model already holds; freed budget flows to fresh facts.
• Client cache — ~/.ultramemory/cache.json (ejected by ultramemory enable; user-editable, chmod 600, LRU-bounded at 500 entries / ~1 MB). It memoizes identical recall queries for 5 minutes (a repeat query makes zero HTTP calls) and tracks each session's seen fact_ids for 24 h. Delete the file to reset; corrupt files are silently rebuilt.

Environment tunables:

Env	Default	Effect
ULTRAMEMORY_CACHE=off	on	kill switch — disables the memo + seen cache entirely
ULTRAMEMORY_PREVIEW=off	on	Hermes prefetch reverts to full (non-preview) recall
ULTRAMEMORY_HOOK_BUDGET	2000	Claude Code hook recall budget in characters
ULTRAMEMORY_HOOK_POLICY_BUDGET	12000	hook injection cap on [COMPANY POLICY] turns (whole policies, never truncated)
ULTRAMEMORY_MIN_CONFIDENCE	low	hook skips injection below this recall confidence

Why UltraMemory

• Deterministic recall-first. "Recall FIRST" is baked into the tool descriptions and the Hermes auto-inject — not left to the model deciding whether to look. The hook makes a deterministic injection attempt before every prompt (fail-open, top matches); paired with the active-recall CLAUDE.md rule for the agent's own mid-reasoning lookups, that trio is the real recall-first guarantee.
• Honest about what it doesn't know. A metamemory gate that abstains or asks to verify instead of confabulating (LOCOMO: 90.2% correctly-abstained).

License

Apache-2.0 (see LICENSE). This is the open-source client surface. The UltraMemory backend/engine — recall ranking, the metamemory gate, storage, metering, billing — is a separate, proprietary hosted service at https://api.ultramemory.us.