Memory

Durable memory infrastructure

Append-only, auditable memory with cryptographic provenance. Track what agents knew, retrieved, and concluded, all with proof.

Benchmark

#1 on LoCoMo

XMDB achieves 93.2% accuracy on LoCoMo, the most widely cited benchmark for AI agent memory.

Leaderboard

1
XMDB
93.2%

Leader

2
MemU
i
92.09%

1.11 pts behind XMDB

3
MemMachine
i
91.69%

1.51 pts behind XMDB

4
Backboard
i
90.1%

3.1 pts behind XMDB

5
Hindsight
i
89.61%

3.59 pts behind XMDB

6
Memvid
i
85.65%

7.55 pts behind XMDB

7
Zep
i
80.32%

12.88 pts behind XMDB

8
PowerMem
i
78.7%

14.5 pts behind XMDB

9
Letta
i
74%

19.2 pts behind XMDB

10
Mem0
i
66.88%

26.32 pts behind XMDB

11
OpenAI (full-context)
i
52.9%

40.3 pts behind XMDB

Leaderboard uses latest publicly posted LoCoMo scores as of February 18, 2026. XMDB is internally verified. Use the i button next to externally sourced systems for notes and links.

XMDB Category Breakdown

97.5%
Multi-hop
275/282
92.8%
Temporal
321 questions
92.3%
Single-hop
841 questions
89.6%
Open-domain
96 questions

Architecture

XMDB uses a two-tier architecture. Simple factual queries are handled by hybrid BM25 + vector retrieval with cross-encoder reranking. Complex queries requiring synthesis across multiple sessions are automatically routed to full-context reasoning. The system uses benchmark-provided question type labels for routing; in production, a lightweight query classifier replaces this.

Hybrid retrievalCross-encoder rerankingStructured claims extractionAdaptive complexity routingMulti-model arbitrationCategory-aware prompting

LoCoMo evaluates long-term conversational memory across 1,540 questions spanning 10 multi-session dialogues. Scoring uses LLM-as-judge evaluation. XMDB is internally verified; external scores are sourced from vendor benchmark pages and papers linked via the info buttons in the leaderboard.

Local-first architecture

XMDB provides durable system memory that allows AI systems to operate over long horizons without losing or silently mutating critical context. It sits beneath applications and agents as a stable memory substrate.

All memory data lives locally in SQLite, giving you full control over your data. The append-only event log is the source of truth, while derived views provide fast access to current state.

SQLite storage

Portable, reliable, no external database dependencies

Works offline

Full functionality without network connectivity

Optional sync

Encrypted sync relay when you need it

Core components

Event Log

Append-only, hash-linked events with device signatures

Compiler

Derives items, claims, edges, and pins from events

Retrieval

Query and pack with policy filtering and attestation

Beliefs

Track agent conclusions with confidence, sources, and temporal validity

Federation

Share memories between agents with cryptographic proof. Receiving agents verify proofs and record why they trusted the data.

event.json
{
  "event_id": "evt_7f3a9b2c",
  "event_type": "capture",
  "workspace_id": "prod-ws",
  "scope_id": "deployments",
  "item_id": "itm_4d8e1f6a",
  "content": "Deployed v2.1.0 to production",
  "source_kind": "deployment",
  "source_ref": "github.com/org/[email protected]",
  "prev_hash": "sha256:8a4f2b...",
  "event_hash": "sha256:c3e91d...",
  "sig": "ed25519:Kx9mQ...",
  "observed_at": "2025-01-07T10:30:00Z"
}

Append-only event log

Every memory write creates an immutable event. Events are hash-linked to form a chain, and optionally signed with Ed25519 device keys. The log can never be silently modified.

  • Per-scope hash chains for integrity verification
  • Device signatures for provenance tracking
  • Tombstones for logical deletes (audit-preserving)
  • Revision history for item updates

Deterministic context packs

The /pack endpoint builds context packages with token budgeting, pin awareness, and cryptographic attestation. Every pack includes a hash and optional signature.

  • pack_hash for content verification
  • Optional Ed25519 signatures
  • Token budgeting respects context window limits
  • Pinned items always included
bash
# Request a context pack
curl -X POST https://api.xmdb.cloud/pack \
  -H "Authorization: Bearer $XMDB_API_TOKEN" \
  -d '{
    "workspace_id": "prod",
    "scope_id": "project",
    "q": "deployment",
    "token_budget": 2000,
    "include_pins": true
  }'
# Response includes attestation
{
  "pack_text": "...",
  "pack_hash": "sha256:abc123...",
  "sig": "ed25519:...",
  "sig_alg": "ed25519",
  "chunks_included": 5,
  "tokens_used": 1847
}

Proof at every layer

Every step of agent cognition is cryptographically verifiable.

Receipts

Every context pack includes a cryptographic receipt: proof of exactly what the agent retrieved.

Beliefs

Record what agents concluded, why, and with what confidence. Query beliefs at any point in time with as_of.

Merkle proofs

Checkpoint events into Merkle trees. Generate inclusion proofs for any event without exposing the full chain.

External anchoring

Anchor Merkle roots to RFC 3161 timestamp authorities. Third-party proof that data existed at a specific time.

belief.json
{
  "belief_id": "bel_8a3f2c1d",
  "receipt_id": "rcp_4e7b9a2f",
  "conclusion": "User prefers dark mode",
  "reasoning": "Based on explicit statement in conversation",
  "confidence": 0.92,
  "source_claim_ids": ["clm_1a2b3c", "clm_4d5e6f"],
  "agent_id": "assistant-v1",
  "valid_from": "2025-01-16T10:30:00Z",
  "valid_to": null,
  "status": "active"
}

Tracked agent beliefs

Beliefs track agent conclusions with provenance. Query at any point in time. Invalidate when wrong.

  • Linked to retrieval receipts for full traceability
  • Confidence scores and reasoning captured
  • Temporal validity with as_of queries
  • Invalidation with full audit trail

Search & retrieval

XMDB supports multiple retrieval strategies to find relevant memory across your history.

  • Full-text + vector search (hybrid retrieval) that combines keyword matching with semantic similarity
  • Optional HNSW vector indexing to scale semantic search with efficient approximate nearest neighbor lookups
  • Deterministic extraction + context receipts for verifiable retrieval with cryptographic attestation

API & CLI

Full HTTP API and command-line interface for all operations

API Endpoints

  • GET/healthzHealth check
  • GET/whoamiToken identity
  • GET/v1/itemsList items
  • POST/v1/itemsCreate item
  • POST/captureStore memory
  • POST/querySearch memory
  • POST/packContext pack
  • GET/items/historyRevisions
  • POST/v1/beliefsCreate belief
  • GET/v1/beliefsList (as_of)
  • POST/beliefs/:id/invalidateInvalidate
  • GET/v1/beliefs/diffChanges
  • POST/v1/checkpointsCheckpoint
  • GET/checkpoints/:id/proofMerkle proof
  • POST/checkpoints/:id/anchorRFC 3161
  • POST/v1/federation/shareShare bundle
  • POST/v1/federation/receiveReceive bundle
  • GET/v1/federation/receiptsFed receipts
  • POST/v1/trust/policiesTrust policy

CLI Commands

  • xmdbctl loginAuthenticate
  • xmdbctl statusCheck connection
  • xmdbctl captureStore memory
  • xmdbctl itemsList items
  • xmdbctl packBuild context pack
  • xmdbctl auditAudit bundles
  • xmdbctl ingestBulk import
  • xmdbctl beliefs createCreate belief
  • xmdbctl beliefs listList beliefs
  • xmdbctl beliefs invalidateInvalidate
  • xmdbctl checkpoints createCheckpoint
  • xmdbctl checkpoints proofMerkle proof
  • xmdbctl checkpoints verifyVerify offline
  • xmdbctl checkpoints anchorRFC 3161
  • xmdbctl federation shareShare bundle
  • xmdbctl federation receiveReceive bundle
  • xmdbctl trust policies createTrust policy
Model Context Protocol

Multi-agent support via MCP

XMDB includes a native MCP server that exposes memory tools to any MCP-compatible client. Run locally via stdio or connect to the remote endpoint at https://mcp.xmdb.cloud/mcp.

Available tools

write_memorysearch_memoryget_memory_packget_item_history

Aliases

xmdb.capturexmdb.searchxmdb.packxmdb.item_historyxmdb.whoamixmdb.items_list
View MCP documentation

Remote MCP endpoint

bash
# Remote MCP endpoint (Streamable HTTP)
https://mcp.xmdb.cloud/mcp
# Requires MCP auth token (separate from API token)
Authorization: Bearer $MCP_AUTH_TOKEN

Local stdio mode (Claude Desktop)

claude_mcp_config.json
{
  "mcpServers": {
    "xmdb": {
      "command": "xmdb-mcp",
      "env": {
        "XMDB_BASE_URL": "https://api.xmdb.cloud",
        "XMDB_API_TOKEN": "your-api-token",
        "XMDB_WORKSPACE_ID": "your-workspace"
      }
    }
  }
}
Local-first
SQLite storage
Append-only
Audit log
Least-privilege
Scoped tokens
MCP
Local + remote
Beliefs
Tracked conclusions
Merkle proofs
Selective verification
RFC 3161
External anchoring

Start building with durable memory

Get started in minutes with the CLI or connect your MCP clients.

Contact UsRead Documentation