VelesDB Memory

Start here — three commands that work

pip install velesdb
curl -O https://raw.githubusercontent.com/cyberlife-coder/VelesDB/main/examples/python/hello_velesdb.py
python hello_velesdb.py

To search your own text instead of hand-written vectors, install the opt-in local adapter (pip install "velesdb[embed-sentence-transformers]") and run hello_velesdb_text.py; its first run downloads all-MiniLM-L6-v2.

Expected output, byte-for-byte (read the script — no server, no embedding model):

Query: "tech"
  score=1.000  Rust 1.89 release notes
  score=0.600  AI-generated jazz: the new wave
  score=0.000  Best ramen in Tokyo

Query: "tech + music"
  score=0.990  AI-generated jazz: the new wave
  score=0.707  Rust 1.89 release notes
  score=0.707  Miles Davis discography

An embedding model determines the vector dimension, while your similarity semantics determine the metric. Both are fixed when a collection is created; to change either, create a new collection and re-index your documents.

Give your agent a persistent memory — three more commands:

cargo install velesdb-memory                                    # the local MCP memory server
claude mcp add velesdb-memory -- ~/.cargo/bin/velesdb-memory    # any MCP client works
curl -L https://github.com/cyberlife-coder/VelesDB/releases/latest/download/velesdb-skills.tar.gz | tar -xz -C ~/.claude/skills/

No Rust toolchain? npm i @wiscale/velesdb-memory-node, or grab a prebuilt .mcpb bundle from the official MCP Registry (io.github.cyberlife-coder/velesdb-memory).

Memory used continuously, not just available: integrations/agent-hooks/ wires five Claude Code hooks — SessionStart/Stop/PreCompact resume and save the working context, PreToolUse requires successful recall before an opted-in repository edit, and PostToolUse both records that recall and compiles an oversized tool result before it enters the transcript. One global install covers every project without enabling the edit guard outside explicitly configured repositories.

One memory shared by several clients (Claude Code, Codex CLI, Claude Desktop, Windsurf, Devin CLI): scripts/install-memory-daemon.sh runs velesdb-memory as a single local daemon — HTTPS by default, with a natively generated local CA.

Cargo (Rust + REST server): cargo install velesdb-server velesdb-cli — Docker (multi-arch linux/amd64 + linux/arm64): docker run -d -p 8080:8080 -v velesdb_data:/data --name velesdb ghcr.io/cyberlife-coder/velesdb:latest, then curl http://localhost:8080/health.

Browser / edge: the WASM build is ~710 KB gzipped and runs entirely client-side (TypeScript SDK). REST: 54 REST endpoints (OpenAPI spec). Full matrix: installation guide.

Why VelesDB

• One database instead of three. Vectors for "what feels similar", a graph for "what is connected", typed columns for "what I know for sure" — normally three deployments, three query languages, and glue code. Here it is one binary and one language.
• A memory that can be audited, not just queried. Every recall can show the evidence behind it; every compression decision carries a rule id, a reason, and a risk level. Deterministic by construction — no model in the write path, so no drift and nothing to re-litigate.
• Local-first is a sovereignty decision, not a latency one. No cloud, no API key, no data processor: air-gapped if you want it, in your jurisdiction by default. Why that matters · positioning in depth.

How it works, in plain terms

Four things happen, and none of them calls an AI provider.

1 · It stores facts, not conversations. You give it one statement — "the API port is 6333 because 3000 collided with the web UI" — and it lands in a local file store. No model call, nothing sent anywhere.

2 · It finds them by meaning. Asking "which port did we settle on" reaches that fact even though none of the words match. A local embedding model turns text into coordinates; close meaning means close coordinates.

3 · It connects them, and that is the part a search engine cannot do. Each fact is linked to the topics it mentions. why() starts from the best match and then walks those links, so it returns the answer plus the facts that explain it — including ones sharing no vocabulary with your question.

The links have to exist. Store facts one by one and the graph stays flat, so why() behaves like a search. Hand a paragraph to remember_extracted and it splits it into facts and wires the links for you.

4 · It compresses what is too big, before you pay for it. Give the compiler your accumulated context and a token budget; it returns a smaller version with one recorded decision per fragment — kept, abstracted, or dropped — and a handle to fetch any original back verbatim. Same input, same bytes out, every time. That is what the 82.5 % below measures.

What no one else combines

1 · Three engines, one query

One statement crosses all three — similarity, relations and typed filters, no glue code:

MATCH (doc:Document)-[:AUTHORED_BY]->(author:Person)
WHERE similarity(doc.embedding, $question) > 0.8
  AND author.department = 'Engineering'
RETURN author.name, doc.title
ORDER BY similarity() DESC LIMIT 5

2 · A memory that shows its evidence — why()

Most "agent memory" is vector recall: it finds text that looks like your query. VelesDB connects memories with typed links, so it can answer why something happened by walking the graph to context that shares no words with your question — across process restarts, offline, no API key:

from velesdb import MemoryService            # pip install velesdb

mem = MemoryService("./agent_memory")        # a real on-disk store; survives restarts
reason = mem.remember("Robert is recovering from knee surgery")
mem.remember("Booked the aisle seat on Robert's flight", links=[(reason, "because")])

# A *new* process, weeks later, reopens the same store and asks why:
mem.why("why the aisle seat on Robert's flight?")   # walks booking → reason — recall() can't

MemoryService defaults to the offline hash embedder: deterministic, but lexical rather than semantic, so unrelated wording can score 0.000. Opening it now says so once on stderr. For meaning-based recall, pass embedder="ollama"; follow Real semantic recall in 5 minutes.

Memories are permanent by default; forget(id) deletes one, ttl_seconds gives a fact a durable expiry. Every remember auto-stamps its storage day, so recency-weighted recall works with zero setup. Same wedge in Python, Node, the MCP server, and in-memory in the TypeScript SDK.

Proof it is not a weak-embedder trick — four runnable demos in which recall stays blind to the reason even under a real semantic embedder (ollama / all-minilm), because the reason is connected by a decision rather than by surface similarity: why_across_sessions.py (survives a process restart) · why_magic_constant.py (a business reason sharing no words with the code) · memory_builds_its_own_graph.py (raw prose in, auto-wired graph out) · why_magic_constant.mjs (Node). Benchmark position, including LoCoMo and why cross-lab scores are not fairly comparable: BENCHMARK.md · Agent Memory guide.

3 · A deterministic context compiler

Agents burn most of their budget re-reading redundant context. compile_context / compile_transcript (MCP, or ContextCompiler in Rust) shrink it with no LLM and no network:

• Deterministic — the same input always compiles to the same bytes, asserted twice per run in every committed benchmark. That also yields a byte-stable cache prefix provider prompt-caching can actually hit.
• Auditable — explain_compilation gives every kept or dropped fragment a stable rule id, a reason and a risk level.
• Reversible — over-budget content becomes a recoverable ctx://source/ handle; retrieve_context_source brings the original bytes back on demand.
• Bounded — it compresses only what your agent explicitly hands it, never the harness's system prompt, and nothing enters recallable memory without an explicit remember.

Code, URLs, numbers and negative constraints survive verbatim. The velesdb-context-optimizer skill teaches the workflow — including when not to compress.

Proof — three numbers, each tied to its harness

No figure here is an estimate from a slide; each links to the log or script in this repo that produced it.

Same campaign, less flattering: 10.9 % on cropped screenshots, 14.7 % on a 36-turn day-scale arc, 15.1 % input tokens on the direct Messages API, and 2.5 % for the no-screenshots variant — that spread is the measured value of the media mechanisms, so we publish it as prominently as the headline. Honest reading, limitations and full protocol. Every number on this page is CI-guarded by a promise contract that pins the README to its committed sources.

Billed A/B sessions (2026-07-19, claude-sonnet-5; raw logs committed verbatim):

* Two turns' grading key was later found defective (both arms scored full marks there; the parity conclusion stands) — disclosure. Over a 36-turn session compiled context grows 1.7× slower, so one session lasts far longer before hitting the window.

Memory retrieval quality, public test sets, no AI grader in the loop: +7.2 pts multi-hop (HotpotQA), +9.7 pts time-scoped recall (TimeQA), +29 pts on a controlled task needing both engines at once — BENCHMARK.md.

End-to-end search (canonical): search p50 450 us (10K, 384D, WAL ON; measured 2026-03-27 on 1.7.2, Balanced then at ef 128 — harness, report) · quantization PQ (2 × dim / m smaller: dim/4 at the default m = 8), RaBitQ (32x smaller), SQ8 (4x smaller), Binary (32x smaller) — scope & caveats.

Index-only micro-benchmarks (no WAL, no payload, hot cache — not comparable to the end-to-end figure above), each reproducible with cargo bench -p velesdb-core --bench <name>: HNSW Search index-only (10K/768D, k=10) 55 us (hnsw_benchmark -- hnsw_search_latency) · SIMD Dot Product (768D, AVX2) 21.7 ns (simd_benchmark) · Recall@10 balanced 99.8% and accurate mode 100% (10K/128D, recall_benchmark) · BM25 Sparse Search index-only (10K docs, top-10) 57.6 us (sparse_benchmark -- top10_10k_corpus).

ef_search is each preset's base value, used as-is for k ≤ 32 on up to 10K vectors; ef_search_for_scale raises it beyond. Recall@10: recall_benchmark, 10K/128D, 2026-09-10 on Apple M5 Pro.

Distance metrics — 5 with SIMD acceleration (AVX-512, AVX2, NEON). At 768D on the i9-14900KF reference (AVX2, hot cache; the April 3, 2026 run in docs/BENCHMARKS.md §1): Cosine 32 ns · Euclidean 26 ns · Dot Product 22 ns · Hamming 35 ns · Jaccard 27 ns.

ColumnStore — typed columnar filtering, 130x faster than JSON scanning at 100K rows on the i9-14900KF reference (JSON scan 3.84 ms → ColumnStore 29.5 us). The ratio is hardware-dependent: on Apple Silicon (M5 Pro, 2026-07-20) the JSON scan itself runs ~2.8× faster, so the same bench measures ~50–105x while the ColumnStore's absolute time holds (~27 µs).

Provenance: Intel Core i9-14900KF (x86_64, AVX2). Per-machine figures vary; Apple-Silicon cross-checks, the SIFT1M standardized ANN run and the full methodology live in docs/BENCHMARKS.md. Reproduce the end-to-end figure with python benchmarks/velesdb_benchmark.py --recall.

Pick your entry point

Tool parity per surface is published honestly — including where a surface is still behind: memory crate README. Worked examples: examples/.

Full API reference: docs/reference/api-reference.md | OpenAPI spec: docs/openapi.yaml | Server security: docs/guides/SERVER_SECURITY.md

How it compares

Sweet spot: vector + graph + structured filtering in one engine, local-first, auditable. Not the best fit (yet): a managed cloud service with a multi-node distributed cluster. Competitor figures are typical public ranges, not a head-to-head run we performed — run your own. Detailed comparison against agent-memory products (Mem0, Zep, Letta), as of mid-2026: docs/WHY_VELESDB.md.

VelesDB Premium — the enterprise control plane

The core engine is source-available and stays that way. Premium adds the company-grade layer on top of the same binary, for organizations running agent fleets on sensitive data: RBAC on every endpoint including the memory and context-compiler surfaces · audit trail (who, what, when — metadata only, GDPR-conscious) with forensic replay · multi-tenancy with hard per-tenant isolation and two-level deletion rights · clustering and air-gapped deployment · a WebAdmin UI for operators.

Pricing on quote — [email protected] · velesdb.com. Built by Wiscale (France; GDPR and data-sovereignty native).

Known limitations — honest boundaries

The items below are deliberate trade-offs or Premium-tracked features, not correctness gaps — the Community Edition is production-ready for single-node, local-first deployments. We publish them next to the strengths, including the ones we have not fixed yet.

Internal technical limitations (query-planner approximations, plan-cache semantics): docs/reference/KNOWN_LIMITATIONS.md.

Contributing & contact

Quality bar: cargo test --workspace — 9k+ tests across Rust, TypeScript and Python run in CI on every merge; exact commands in QUALITY_BAR.md.

Contributions welcome — start with CONTRIBUTING.md and the good first issues. Security reports: SECURITY.md. Roadmap: ROADMAP.md · Changelog · DeepWiki.

License: VelesDB Core License 1.0 (source-available). Premium: commercial license. Contact: [email protected] · velesdb.com

The name nods to Veles, a deity of old Slavic myth — a keeper of hidden knowledge and boundaries.