Metadata-Version: 2.4
Name: ultramemory-hermes
Version: 1.7.0
Summary: UltraMemory — self-learning, metamemory-gated long-term memory provider for Hermes Agent (and any MCP client). One API key = your own private tenant.
Project-URL: Homepage, https://ultramemory.us
Project-URL: Documentation, https://ultramemory.us
Author-email: UltraMemory <support@ultramemory.us>
License: Apache-2.0
License-File: LICENSE
License-File: NOTICE
Keywords: agent,hermes,llm,mcp,memory,metamemory,rag
Classifier: Intended Audience :: Developers
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3 :: Only
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Requires-Python: >=3.9
Requires-Dist: httpx
Description-Content-Type: text/markdown

![UltraMemory](assets/ultramemory-icon-512.png)

# UltraMemory — cross-tool memory for your AI

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

[![PyPI](https://img.shields.io/pypi/v/ultramemory-hermes?color=2dba4e&label=pypi%3A%20ultramemory-hermes)](https://pypi.org/project/ultramemory-hermes/)
[![License](https://img.shields.io/badge/license-Apache--2.0-blue.svg)](LICENSE)
[![MCP](https://img.shields.io/badge/MCP-Streamable%20HTTP-7c3aed.svg)](https://modelcontextprotocol.io)

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

```bash
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).

## Tools

The MCP server (`https://api.ultramemory.us/mcp`, Streamable HTTP) exposes six 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. |
| `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. |
| `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):**
```bash
claude mcp add --transport http ultramemory https://api.ultramemory.us/mcp \
  --header "Authorization: Bearer um_YOUR_KEY"
```

**Cursor / generic `mcp.json`:**
```json
{ "mcpServers": { "ultramemory": {
  "url": "https://api.ultramemory.us/mcp",
  "headers": { "Authorization": "Bearer um_YOUR_KEY" }
}}}
```

**Claude Desktop (mcp-remote bridge):**
```json
{ "mcpServers": { "ultramemory": {
  "command": "npx",
  "args": ["mcp-remote@latest", "https://api.ultramemory.us/mcp",
           "--header", "Authorization: Bearer um_YOUR_KEY"]
}}}
```

**Hermes:**
```bash
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:**
```bash
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 auto-discovery** — no API key needed. Add
this `mcpServers` block to `~/.gemini/settings.json` (`httpUrl` is Gemini CLI's streamable-HTTP
transport; on first use the CLI detects the 401, discovers the OAuth endpoints, performs dynamic
client registration, and opens the sign-in flow automatically):

```json
{
  "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):

```markdown
## 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`:

```bash
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`](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`](hooks/)) — 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](skills/ultramemory-snapshot/)
  (Claude Code ≥ 2.1.163).

Both are fail-open and copy-paste runnable. See [`hooks/README.md`](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_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. Recall-first, guaranteed.
- **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](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`.
