memory & context for AI agents · open source · local-first

Load only the context your agents actually need.

Perseus is the memory and context layer for AI agents. It cuts the token bill and makes agents smarter by handing them the exact facts a task needs.

Not the whole notebook, re-read from page one on every single call.

Try the live Vault demo ↗ Estimate your savings Explore products
611×
warm vs cold renders
measured on the Perseus repo itself
3 → 1
listings per tool advertised
~67% smaller tool payload
73.8%
LongMemEval QA accuracy
official harness · vs Zep 63.8%
87.9%
LOCOMO accuracy
vs Mem0 82.2%, Zep 33.8 · mem0's own harness
$ pip install perseus-ctx · MIT · zero telemetry
The problem

Agents re-read everything, every call. You pay per word.

An agent with no memory layer reloads its entire working set on every turn: all its tools, the whole history, every doc, and the model is billed for every token, every time.

Perseus resolves the workspace into verified facts before the context window opens, and retrieves only the memory a task needs. It hands the agent the one page it's looking for.

avg tokens per task, LongMemEval_s (n=60, same questions both arms)
Full-context baseline104,180
With Perseus + Vault25,497
measured · provider-billed tokens, Plutus ledger · accuracy 55.0% → 66.7% on the same run · full-context is a no-memory-management upper bound
Why pay for it

Three reasons the memory layer earns its place.

It pays for itself in tokens.

Resolve-before-context and a slimmer tool manifest mean the model reads a fraction of the tokens on every call.

  • 611×warm-cache renders (619.1s → 1.014s, 22,528 directives)
  • 67%smaller tool schema (each tool advertised once, not 3× aliased)
  • 1.0semantic equivalence (20/20 A/B pairs): same answers, fewer tokens

It unlocks deployments others can't.

Fully local-first and air-gap ready: the encrypted, offline posture regulated and sovereign buyers require.

  • Runs offline · zero telemetry · no cloud calls
  • AES-256-GCM encryption at rest
  • Immutable, crypto-chained audit journal

It makes agents smarter.

Durable memory means agents remember decisions, conventions, and past failures, so they stop repeating them.

  • 73.8%LongMemEval QA (vs Zep 63.8%, Mem0 49.0%)
  • Bi-temporal history: what was true & when
  • Hybrid recall · confidence decay · dedup
Value calculator

Plug in your fleet. See the token bill.

Illustrative economic model, not a measured result for your workload.
Agents in your fleet100
5500
Calls per agent / day1,000
1005,000
Model tier / 1M input tokens · FedRAMP-authorized platforms, Jul 2026 list
How this is estimated. We apply the measured 75% token reduction from our cost-savings certification (LongMemEval_s, n=60 stratified: 6.25M → 1.53M provider-billed tokens for the same question set, accuracy 55.0% → 66.7%, dollars read back from a Plutus ledger) to the task volume you set (agents × calls/day × 365). The baseline arm is full-context stuffing, a no-memory-management upper bound: if you already do retrieval, your reduction will be smaller. Editing the baseline scales both arms by the measured ratio.

Model tiers are illustrative input-token list rates (Jul 2026) for the platforms the public sector actually buys: Frontier ≈ Claude Opus on Bedrock ($5); Standard ≈ GPT-4o on Azure / Claude Sonnet ($2.50–$3); Efficient ≈ Claude Haiku / Gemini Flash ($1–$1.50); Economy ≈ GPT-5-nano / Amazon Nova / Gemini Flash-Lite ($0.05–$0.10). The GovCloud toggle adds a flat 25% for AWS GovCloud / Azure Government / Gemini for Government regions. Every field is editable, so plug in your own contracted rate.

Counts input/prompt tokens only (not output); each currency is shown at its own list price, with no exchange-rate conversion. This is an illustrative model, not a measured result for your workload.

Prices assessed July 2026 against AWS Bedrock (incl. GovCloud), Azure OpenAI / Azure Government, and Google Vertex / Gemini for Government. Model versions and promotional rates move quickly (e.g. Claude Sonnet 5's launch rate expires Aug 31, 2026), so treat the presets as anchors, not contract quotes, and enter your own contracted rate.
Estimated annual token savings
$0
$0 / month · 0 tokens saved / yr
Baseline spend / yr$0
With Perseus / yr$0
Token savings only. Excludes output tokens, infra, and the compounding value of fewer agent errors.
Two products, one layer

The context engine and the memory store.

Perseus

v1.0.24

The live context engine. Resolves git, services, tests, and workspace state into verified facts before the session opens, so agents never burn turns re-orienting.

  • 611×warm-cache renders, measured on this repo
  • 33MCP tools resolving live state at call time
  • MITlocal-first · zero telemetry · air-gap ready
$pip install perseus-ctx
Explore the context engine →

Perseus Vault

v2.20.2

Persistent, encrypted memory in a single Rust binary. One file, no Docker, no Postgres, no cloud: durable agent memory that works with any MCP host.

  • 55+MCP tools (advertised once per tool, not 3× aliased)
  • 73.8%LongMemEval QA · hybrid BM25 + dense recall
  • AES256-GCM at rest · bi-temporal · audit journal
$curl -sSf …/install.sh | sh
2026 paradigm

The memory layer of the AI harness.

2026's consensus: the model isn't the product — the system around the model is. Mitchell Hashimoto coined it. OpenAI proved it (1M lines, zero hand-written). Anthropic published the blueprint. The new discipline is harness engineering — and memory is the layer every harness needs.

Context Engineering

Stage 3 → 4

Perseus resolves live workspace state into verified facts before the session opens. Agents start with exactly the context they need — not an encyclopedia.

Knowledge Loops

Stage 3 → 4

Vault's remember/recall/reflect turns agent failures into systematic feedback. "Every time an agent makes a mistake, engineer a solution so it never makes that mistake again."

Security & Trust

Stage 4

AES-256-GCM encrypted at rest. Local-first, zero API keys, zero cloud dependencies. Scoped, auditable agent access — the harness trust boundary.

Read the full harness brief → Agent = Model + Harness
Sovereignty & compliance

Built for the deployments the cloud can't reach.

Perseus runs entirely on your machine. No data leaves the boundary, nothing phones home, and there is no service to depend on, which is exactly what regulated, classified, and sovereign environments require.

Open formats, open source, no vendor lock-in.

Read the full sovereignty brief →
Air-gapped
Fully offline. Zero cloud dependencies, zero telemetry.
Encrypted
AES-256-GCM at rest with immutable, chained audit trail.
MIT licensed
No copyleft, no GPL/AGPL, no lock-in. Yours to run.
Government-ready
SBOM per EO 14028 · NIST AI RMF aligned · EAR99.

Measured, reproducible.

Every figure below is measured on named hardware with a rerunnable script. Economic models are labeled illustrative; benchmarks are not.

GitHub README lead
See both README leads →
Perseus-Computing-LLC / perseus · README.md

Perseus

The memory & context layer for AI agents: load only the context they actually need.

Your agents re-read their whole notebook from page one on every call, and you're billed per word. Perseus hands them just the page they need: it resolves live workspace state into verified facts before the context window opens, and pairs with Perseus Vault for durable, encrypted memory. The result: 73.8% on LongMemEval, a 67% smaller tool schema, and 611× warmer renders. Local-first, air-gap ready, MIT.

pip install perseus-ctx MIT patent pending