MEMORY.
AI.
Graph-native memory for AI agents. Persistent context, temporal versioning, adaptive retrieval, and temporal versioning — all in one API call. Your agents never start from zero.
Live memory console
Change your mind.
Watch memory keep both answers.
Three sessions go into a real HydraDB graph-node. Session 5 contradicts session 0. Nothing is deleted — the old fact is closed and linked by SUPERSEDES, so “what’s true now” and “what changed” are two reads of the same data. The last probe asks something never stated, and the honest answer is nothing at all.
Retrieved subgraph
Ingest the sessions, then run a probe.
Trace
Memory infrastructure
built for production.
Not a wrapper around vector search. A full memory operating system with graph storage, temporal versioning, and honest abstention baked in.
Graph-Native Memory
Store memories as a knowledge graph with entities, relationships, and temporal edges — not flat embeddings.
Temporal Versioning
A revised fact never overwrites the old one. The old fact is closed with valid_to and linked by SUPERSEDES, so history stays queryable.
Honest Abstention
A Cypher MATCH that returns zero rows ends the request. There is no nearest neighbour to fall back on, so "not in memory" is reachable.
BYOK Security
Bring your own keys. Data never leaves your infra. Full tenant isolation with SOC 2 in progress.
No SDK To Adopt
Three JSON endpoints over HTTP — ingest a session, ask a question, read the stats. Nothing to install, no framework to marry, no embedding service in the loop.
Auditable Answers
Every fact keeps an ASSERTS edge to the exact turn that stated it, so an answer can quote its own source instead of asserting it.
Four steps.
Zero overhead.
Connect & Deposit
Install the SDK in one line. Pass your API key. Every agent call automatically routes through the Ledger memory layer — zero config required.
Write Memories
Claude extracts (subject, predicate, object) facts per turn. Each is written as a Fact vertex with a validity interval, plus ASSERTS and ABOUT edges — via UNWIND batches, the only write form HydraDB executes.
Retrieve Intelligently
A question is planned into a bounded lookup: current truth, full history, or an algo.SSpaths traversal from a named entity. No embeddings, no reranking — if it returns nothing, the answer is "not in memory".
Update Without Forgetting
A contradicting statement closes the old fact rather than replacing it. "What do you believe now" and "what did you believe in March" become the same query with a different time filter.
We don't just claim #1.
We publish the numbers.
Scored on LongMemEval, locally, with the run and its limits stated rather than rounded off. Updated monthly. Methodology is open.
ledger: 116 instances of the oracle split across all six question types, qwen3.5:4b local, judged by qwen2.5:7b · others: overall LongMemEval-S as reported by their authors, on a harder split with frontier models · reproduce with scripts/run-eval.mjs
Simple Storage Pricing
No per-seat fees. No API limits. Pay for what your agents actually use.
For developers exploring Ledger
For agents scaling in production
For enterprise-grade deployments
BYOC · Air-gapped · Custom SLAs · Dedicated infra · Forward-deployed engineer
Your agents.
Never forget.
Connect your first agent in under two minutes, against a graph-node running on your own machine. The memory layer your agents actually deserve.