WritersLogic Holographic Memory

Local, encrypted, provenance-aware memory for AI agents.

An embeddable memory and search engine that runs on your machine: hybrid document search and a vector-symbolic fact store in one Rust library, with optional encryption at rest and signed provenance. No service calls, no telemetry.

Visit holographic-memory.comView on GitHub

Open source and local-first. WritersLogic Holographic Memory Cloud, for encrypted sync, backup and verification, is in development.

Why it exists

Agent memory is mostly sold as a hosted API or a database server you run and secure yourself. Holographic Memory puts the memory inside your application, with encryption, provenance and erasure in the open-source engine.

  • Local first

    A library you link into your agent, not a service you call. No network calls, no telemetry.

  • Encryption in the open engine

    AES-256-GCM at rest with an Argon2id-derived key, in the Apache-2.0 engine rather than an enterprise tier.

  • Provenance you can verify

    A signed audit log and COSE, SCITT and C2PA statements over Ed25519 with W3C DID identity.

  • Purge that erases

    Delete removes a record from every query at once. Purge rewrites the journal without it and overwrites the old generation with zeros before it returns; a test scans every store file for the deleted text.

Measured, with sources

  • SciFact retrieval quality, nDCG@10

    0.729 hybrid

    vs HMS BM25 alone 0.688, HMS dense alone 0.645. 5,183 documents, 300 queries, MiniLM embeddings supplied by the harness. Quality only: the run was on a loaded machine, so no latency is quoted.

    benchmarks/results/public_scifact.json

  • Lexical search throughput vs SQLite FTS5

    4.48x the queries per second

    vs SQLite FTS5 (3.53.4); HMS p99 0.52 ms vs 5.13 ms; HMS peak memory 1.07x SQLite's. NFCorpus, 324 queries x 8 passes, 11 interleaved rounds, both engines driven from Node 26 on one Intel Core Ultra 9 185H host (WSL2). One host, one workload.

    benchmarks/results/serving_memory_v40.json

On one Intel Core Ultra 9 185H machine, vector search served 2.29x the queries per second of hnswlib on SIFT-1M at recall 0.95, and text search 2.21-2.76x tantivy on three BEIR datasets at the same nDCG@10. Index build time, serving memory and text cold start are still behind. The charts, sources and losses are on holographic-memory.com.

Everything else lives on the product site

The source on main is Apache-2.0. Packages published so far (crate 0.6.1, holographic-sdk 0.1.2 on PyPI) were released under AGPL-3.0-or-later and stay available under those terms.