Universal AI Memory

Local-first memory with encrypted payloads and authenticated search metadata.

A Python reference service that exposes API and MCP entry points while keeping storage, embedding, entity extraction and authorization behind injected adapters.

Python 3.12 · FastAPI · Pydantic · MCP  ·  store · retrieve · search · forget

01 / SYSTEM CARD

• Primary boundary: bearer capability → API / MCP → manager → adapters / vector store

• Documentation stance: architecture and presentation only; application source is intentionally untouched.

• Visual language: Y2K terminal grid, brutalist labels, maximalist contrast, readable static SVG fallback.

Related MCP server: Memory MCP

02 / ARCHITECTURE

The map above is a compact view of the repository's visible runtime boundary. It is designed to answer three questions quickly: where input enters, where policy or domain logic lives, and where external side effects leave the system.

03 / DESIGN NOTES

• Private payload fields are encrypted; backend-visible vectors, IDs, timestamps and blind indexes remain separate concerns.

• The included backend and deterministic embedder are development implementations, as documented by the project.

This upgrade adds documentation assets only. It does not change source files, dependency manifests, tests, workflows or runtime configuration.

04 / EXISTING PROJECT DOCUMENTATION

Universal AI Memory starter

Python 3.12 local-first reference service with injected embedding, entity extraction, authorization, and vector persistence adapters. Private payload fields use AES-256-GCM; normalized entity labels are represented in the backend as HMAC-SHA-256 blind indexes. The included backend and deterministic embedder are development implementations.

Install and test

python3.12 -m venv .venv
. .venv/bin/activate
pip install -e '.[test]'
pytest

Tests are in-process and make no network requests. openapi.yaml is included in the wheel as universal_memory/openapi.yaml.

Run the HTTP API

Provide two distinct base64 keys and a non-empty bearer API key. Key-generation commands print secrets to the terminal; store real values in a secret manager, not source or logs.

export UNIVERSAL_MEMORY_ENCRYPTION_KEY="$(openssl rand -base64 32)"
export UNIVERSAL_MEMORY_INDEX_KEY="$(openssl rand -base64 32)"
export UNIVERSAL_MEMORY_API_KEY="$(openssl rand -hex 32)"
universal-memory-api

The API binds to 127.0.0.1:8000. Every operation requires Authorization: Bearer $UNIVERSAL_MEMORY_API_KEY. Runnable mode grants that key read, write, and delete capabilities. Embedded applications must inject both a manager and an AuthorizationPolicy into create_app; omitting a policy denies all requests. A recognized credential without the operation's capability receives 403, while a missing or invalid credential receives 401.

curl -H "Authorization: Bearer $UNIVERSAL_MEMORY_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{"content":"Met Alice at launch","metadata":{"project":"demo"}}' \
  http://127.0.0.1:8000/store

Run MCP

MCP tools are independently gated as read, write, and delete. No capability is granted by default. The local stdio launcher accepts an explicit comma-separated capability set:

export UNIVERSAL_MEMORY_MCP_CAPABILITIES=read,write,delete
universal-memory-mcp

Use create_mcp_server(manager, capabilities={Capability.READ}) for a least-privilege embedded server. An MCP transport with per-client authentication should construct separately scoped sessions rather than sharing ambient capabilities.

Data and validation model

A store request may provide content, strict-JSON metadata, an encrypted source, a timezone-aware event_timestamp, and an optional ID. created_at is always service generated. Event chronology uses event_timestamp. Entity labels extracted during storage are encrypted in the payload and exposed only to authorized responses; filters use keyed blind indexes.

VectorStoreManager is the enforcement boundary even when HTTP/MCP schemas are bypassed. Default caps are: content 65,536 UTF-8 bytes, source 2,048 bytes, ID 256 bytes, canonical metadata 65,536 bytes and depth 12, 64 entities of 256 bytes each, 4,096 vector dimensions, and 10,000 records. Metadata must be acyclic strict JSON with string object keys and finite numbers. Validation and canonicalization occur before embedding. Limits can be reduced through MemoryLimits.

ID and entity selectors cannot be combined. Delete-by-entity authenticates complete records before trusting blind indexes or IDs. Malformed, unauthenticated, or corrupt records produce a sanitized domain error and are not used for sorting, scoring, or deletion.

Security boundaries

• Content, metadata, source, event timestamp, and raw entity labels are encrypted in the payload.

• Vectors are not encrypted and may leak semantic information. IDs, ingestion/event timestamps, nonce, ciphertext length, vectors, and HMAC entity indexes are visible to the backend.

• Clear created/event timestamps plus digests of the vector and entity-index set are encoded canonically into AES-GCM associated data. This authenticates those fields; it does not make them private.

• Blind indexes hide raw labels but permit equality correlation. Use an independent high-entropy index key.

• The API uses bearer authentication, not transport encryption. Keep the default loopback binding or add trusted TLS before crossing a network boundary.

• The in-memory backend loses records on restart and is not suitable for multi-process or production deployment. Production adapters also need durable atomic capacity enforcement, key rotation, auditing, and operational controls.

• The service does not intentionally log plaintext or secrets. Decrypted API/MCP responses are returned only after capability checks.