Vault Agent Memory
Website · Founding 10 challenge · Live evidence · Reproduce independently · Methodology · Install
Vault is a local-first, backend-agnostic memory governance layer for AI agents.
Vault Agent Memory gives Codex, Claude Code, Hermes, OpenClaw, n8n, Coze, and other agents one governed memory vault to share. It is not trying to be another notes app or vector database. It helps agents decide what should be remembered, who can use it, whether it is still current, and how to roll it back when it is wrong.
Vault's product boundary is the governance contract, not a specific backend. The same candidate-first review model can run on local SQLite, a self-hosted central memory host, a Supabase cloud adapter, or a future managed Vault Cloud backend.
Vault works standalone, and it can also become the governed memory backend for other agent memory frameworks. You can use Vault by itself with CLI, MCP, Gateway, and local SQLite; or let frameworks such as Hermes, OpenClaw, Letta, mem0, Claude Code, and Codex connect through Vault APIs while Vault keeps the review, audit, lifecycle, and backend boundaries consistent.
The core multi-agent model is single-host sharing, multi-host governed sync: agents on one trusted machine can share the same local Vault, while agents on other machines or hosted runtimes can read approved memory and submit candidates. Only a trusted sync host promotes official memory, runs lifecycle jobs, and pushes reviewed read copies back out.
The Python package and existing install path remain vault-for-llm.
Vault is for people already building or working with agents. The main interface should still not be a long CLI manual: ask an agent to install Vault, answer a few setup questions, then read a short daily memory report.
New here? Start with the public website. It explains why Vault is a memory foundation—not merely another memory system—and publishes paired A vs A + Vault evidence with the full method and raw artifacts.
Release Maturity
Vault Agent Memory v0.10.2 is a Public Beta / Developer Preview release. The local-first path is ready for agent developers to try in real projects, and the multi-agent model is intentionally conservative:
- local SQLite/Markdown remains the source of truth;
- same-machine agents can share one trusted Vault;
- remote or hosted agents read approved memory and submit candidates;
- only a trusted sync host should hold service-role credentials, promote official memory, run Dream/forgetting/archive jobs, and push reviewed read copies;
- central vector search is a derived read layer over reviewed safe summaries, not a second active-memory database.
- backend adapters can change, but the Vault Governance Contract stays the same: approved reads, candidate submissions, review, promotion, audit, and daily reports.
The package is pre-1.0. Advanced Supabase, Gateway, central vector, automation, and benchmark surfaces are useful for developer preview deployments, but should be enabled deliberately and monitored with the generated reports.
Deployment is intentionally layered:
Agents / Apps / Memory Frameworks
-> Vault Memory API plus MCP / Gateway / OpenAPI adapters
-> Vault Gateway / Vault Governance Contract
-> Memory Provider Interface / Backend adapter
-> Local SQLite / Self-host central host / Supabase / future Vault Cloud
30-Second Version
Vault Agent Memory exists because agent memory fails in practical ways:
- a new session acts like it joined the project on day one
- bug fixes stay buried in chat history
- old notes outrank newer decisions
- private observations leak into shared project memory
- remote agents can pollute shared memory if every host can write active truth
- teams cannot tell which memory was reviewed, trusted, or deprecated
flowchart TB
subgraph Agents["🤖 Your Agents"]
C[Codex]
CL[Claude Code]
H[Hermes]
CO[Coze]
N[n8n]
end
subgraph Vault["🔐 Vault — Governed Memory Layer"]
direction TB
Pipeline["Review Pipeline<br/>Privacy · Duplicates · Quality · Source"]
Report["Daily Report<br/>Auto-promote low-risk · Review high-risk"]
subgraph Layers["Memory Layers"]
L0["L0 Bootstrap"]
L1["L1 Rules"]
L2["L2 Context"]
L3["L3 Knowledge"]
end
Ledger["Task Ledger<br/>Live workbench · Handoffs"]
Storage["SQLite / Markdown<br/>Local-first · Zero dependencies"]
end
subgraph Adapters["🔌 Adapters / Backends"]
Obs[Obsidian Sync]
Sup[Supabase]
GW[Gateway API]
SH[Self-host Host]
VC[Vault Cloud Future]
end
Agents -->|propose| Pipeline
Agents -->|search / bounded read| Layers
Agents -->|update status| Ledger
Pipeline --> Report
Report --> Layers
Layers <--> Ledger
Layers --> Storage
Adapters <-->|import / export / sync / managed backend| Vault
style Agents fill:#e1f5fe,stroke:#0288d1
style Vault fill:#f3e5f5,stroke:#7b1fa2
style Adapters fill:#e8f5e9,stroke:#388e3c
Why Vault?
The core workflow is:
propose -> review -> promote -> search -> bounded read -> rollback -> audit
In plain language:
Vault is not about helping agents remember everything. It is about helping teams govern what agents remember, trust, share, forget, and roll back.
Who Are You? Start Here 👇
For Agent Builders: Ask Your Agent To Install It
Copy this prompt into an agent that can run local commands:
Install Vault Agent Memory for this project. Use vault-for-llm[mcp]==0.10.2.
Use the agent-assisted governed-auto memory mode.
Do not show advanced CLI flags first. Ask me only four questions:
1. Which language should Vault use: Traditional Chinese, Simplified Chinese, or English?
2. Should this be an independent vault or a shared vault for multiple agents?
3. Should Vault connect to Obsidian, Supabase, both, or neither?
4. What time should the daily memory report run?
After setup, run a smoke check and tell me:
- where the vault lives
- how I read the daily memory report
- where the local GUI or next action is
Daily rule:
safe, low-risk, sourced memories can be kept automatically;
uncertain, sensitive, conflicting, or strategic memories should go into the
daily report for my review.
The agent will usually run:
python3 -m venv .venv
source .venv/bin/activate
pip install "vault-for-llm[mcp]==0.10.2"
vault quickstart
You can also print the install prompt from Vault itself:
vault guide --intent install
To check for a newer release without changing the environment:
vault upgrade --check
Vault detects common pip, pipx, uv tool, and editable installs and prints the
appropriate upgrade command. This first-stage command is check-only: it does
not install packages or migrate vault.db.
vault quickstart is the small first-run wizard. It asks only for language,
independent/shared memory, optional Obsidian/Supabase connections, and daily
report time. See docs/quickstart.md for the 5-minute
walkthrough and FAQ. Advanced integration flags stay under vault setup-agent.
Agent-assisted quickstart uses governed-auto by default. Internally this is
still the consumer setup path, but that does not mean Vault is a zero-learning
consumer app. Low-risk, sourced candidates that pass privacy, duplicate,
metadata, and quality gates may enter the active vault. Strategy, private,
sensitive, conflicting, or low-trust memories stay in the daily report for
human review. Nothing is hard-deleted automatically.
Daily Use
The intended human surface is small:
- Agents propose reusable lessons while they work.
- Vault checks privacy, duplicates, quality, and source evidence.
- Safe low-risk memories can enter the vault.
- Uncertain decisions are summarized in a daily report.
- The user approves, rejects, defers, or keeps both sides for conflicts.
The report should answer:
- What did Vault remember today?
- What few memory decisions need my attention?
- Are there stale, sensitive, conflicting, or low-quality memories to review?
That is the product shape: more automatic over time, but still governed.
What Vault Is Not
Vault is not an Obsidian replacement.
Obsidian is great for humans reading notes. Vault helps agents use those notes safely, with source ranges and review boundaries.
Vault is not just RAG.
RAG usually focuses on retrieving context. Vault focuses on the memory lifecycle: who wrote it, whether it was reviewed, which agents can read it, when it stops being current, and how to roll it back.
Vault is not a raw chat-history landfill.
It is candidate-first. Agents can suggest memory, but long-term memory should stay source-backed, reviewable, and clean.
Vault is not a zero-setup app-store product for people who do not use agents.
The first public audience is agent-assisted builders: people using Codex, Claude Code, Hermes, OpenClaw, n8n, Coze, or similar systems who want one governed memory layer without studying every internal command.
How Vault Compares
Vault is complementary to memory services and agent runtimes, not a drop-in replacement for all of them.
If you need a managed memory API for an app, Mem0 may be the faster starting point. If you want a single stateful agent that owns its memory and grows over time, Letta may be the better fit. If you already use several agents and want a local, reviewable, auditable memory foundation between them, Vault is designed for that layer.
That difference matters most across machines: Vault treats remote agents as approved-memory readers and candidate submitters, not as direct writers to the official shared memory.
Benchmarks & Verification
Vault does not currently claim a LoCoMo/LongMemEval-style leaderboard score. Those benchmarks measure end-to-end memory QA and should only be published for Vault after running the same benchmark harness under comparable conditions. Vault now has a neutral external-memory comparison harness for retrieval-only runs across Vault, mem0, Letta, and compatible run artifacts; treat current numbers as source-hit evidence recall unless a fixed reader and judge are explicitly included.
What Vault does publish today is reproducible verification for its own product contract:
- Search QA: source-hit, MRR, no-result, and citation-boundary checks for retrieval evidence, not final answer quality.
- Memory-foundation benchmark:
paired external-engine
Aversus engine-plus-VaultA+BValid Recall, forbidden-exposure, latency, and cost deltas, plus a fixed-clock dynamic governance suite. Synthetic contract evidence is labelled separately from full provider-product claims. A retrieval-only VaultGovBench bundle with five distinct clean-state repeats per track for Vault keyword retrieval and mem0 2.0.12 controlled raw insertion is published underbenchmarks/results/vaultgovbench-retrieval-v0.1/89b9156. - README command smoke: public README commands are exercised so install and quickstart examples do not drift.
- Release gates: pytest, release parity, public-boundary checks, package build checks, and clean-install smoke before publishing.
- Integration smoke: local MCP/CLI paths plus optional Supabase, Gateway, OpenClaw, and Coze read-only hosted-reader checks when those integrations are part of the release scope.
- Multi-host governed sync smoke: when remote sharing changes, verify that anon/scoped agents can submit candidates but cannot write active memory or derived indexes, and that the trusted sync host can pull, review, promote, report, and push approved read copies.
See Search QA benchmarking and External memory benchmarks for the current retrieval-only evidence model. The README claim matrix tracks which public claims are positioning, product-contract checks, or implemented runtime capabilities.
Killer Demo: Shared Governed Memory
Run the local demo:
vault demo agent-governance --json
It simulates Codex, Claude Code, and Hermes sharing one governed vault:
- one agent proposes a lesson from a bug fix
- the memory stays a candidate until reviewed
- a reviewer promotes it with source evidence
- another agent finds it later with search and bounded read
- the memory can be deprecated or rolled back when it becomes outdated
The generated demo pack also includes three follow-up guides:
consumer-mode-demo.md, automation-mode-demo.md, and
multi-host-sync-demo.md.
Start here:
- Agents Need Memory Governance, Not Just RAG
- Three-agent shared-memory runbook
- Demo pack
- Strategy docs
3-Minute Demo (Coming Soon)
🎬 A 3-minute walkthrough GIF is coming soon.
In the meantime, here's what it will show:
- Install —
pip install vault-for-llm[mcp]andvault quickstart- Configure — Answer 4 simple questions (language, vault type, integrations, report time)
- Propose — An agent suggests a memory with
vault_memory_propose- Review — The daily report surfaces candidates for human approval
- Promote & Search — Approved memory shows up in
vault_searchwith bounded readsPrefer a text walkthrough? → 5-Minute Quickstart
One-Click Install
macOS / Linux
curl -sSL https://raw.githubusercontent.com/zycaskevin/Vault-Agent-Memory/main/scripts/install.sh | bash
Windows (PowerShell)
irm https://raw.githubusercontent.com/zycaskevin/Vault-Agent-Memory/main/scripts/install.ps1 | iex
After the installer finishes, run vault quickstart to complete setup.
These raw GitHub URLs are available from main and install the pinned release
version used by this README. If you prefer release-tagged documentation, use
the next release that includes these installer scripts.
Source: scripts/install.sh · scripts/install.ps1
Developer Quickstart
pip install "vault-for-llm[mcp]==0.10.2"
vault init ~/Vaults/demo
vault add "First lesson" \
--content "The bug was caused by a missing cache key. The fix was adding provider metadata." \
--project-dir ~/Vaults/demo
vault compile --project-dir ~/Vaults/demo --no-embed
vault search "cache key" --project-dir ~/Vaults/demo
vault --project-dir ~/Vaults/demo map build
vault --project-dir ~/Vaults/demo map read 1 --lines 1-20
vault --project-dir ~/Vaults/demo gui
vault add takes content through --content or --file. For bounded source
reads, use vault map read <knowledge_id> --lines START-END.
For MCP-capable runtimes:
vault-mcp --project-dir ~/Vaults/demo --tool-profile core
Start most agents with core:
vault_searchvault_read_rangevault_memory_proposevault_statsvault_update_statusvault_automation_handoff
Use larger MCP profiles only when needed:
Detailed MCP docs:
Memory Model
Vault uses L0-L3 for memory depth:
New projects create L0-bootstrap. Existing L0-identity paths remain
read-compatible and are never renamed automatically. L0 is startup/project
context, not a Vault-owned identity or human model.
Task Ledger is not L2. It is the live workbench for blockers, next actions, evidence links, due dates, and handoff notes. Only durable lessons, decisions, and summaries should be promoted into L2/L3 after review.
Access is not controlled by layer alone. Use governance metadata:
scope: private, project, shared, publicsensitivity: low, medium, high, restrictedowner_agentallowed_agentsmemory_typeexpires_atvalid_from/valid_untilsupersedes_id
Temporal fact windows are separate from expiry. expires_at means "move this
out of normal recall later." valid_until means "this fact stopped being true,
but keep it for history and audit."
vault memory temporal status
vault memory temporal list --state past
vault search "office location" --exclude-expired
More detail: docs/memory_governance.md.
Automation And Daily Reports
Automation is report-first by default. It can rank candidates, summarize stale memory, suggest consolidation, and prepare a short review queue without silently rewriting long-term memory.
vault daily-report --language en
vault automation brief --pretty
vault automation review-summary --write-summary
vault automation handoff
Enable stronger automation deliberately:
vault setup-agent \
--automation-schedule cron \
--automation-apply \
--automation-auto-promote-low-risk
That path can promote only low-risk, sourced candidates that pass the normal gates, with a per-run cap. Private, high-sensitivity, duplicate, weak, or sourceless candidates stay in review.
Automation docs:
Integrations
Start here:
Obsidian
Import an existing Obsidian vault:
vault import obsidian --vault ~/Documents/ObsidianVault --project-dir ~/Vaults/my-project --dry-run
vault import obsidian --vault ~/Documents/ObsidianVault --project-dir ~/Vaults/my-project --compile
Export reviewed Vault knowledge back into Obsidian-readable notes:
vault export obsidian --project-dir ~/Vaults/my-project --vault ~/Documents/ObsidianVault --dry-run --json
vault export obsidian --project-dir ~/Vaults/my-project --vault ~/Documents/ObsidianVault
The conflict inbox uses explicit resolver choices: accept Obsidian, accept Vault, or keep both.
Deployment Modes And Remote Sharing
Local SQLite remains the simplest source of truth, but Vault does not require one backend forever. Choose the deployment mode that fits the trust boundary:
The governance rules do not change between modes: remote writes enter as candidates, approved memory is the read surface, and a trusted reviewer or policy gate promotes durable memory.
Supabase is an optional cloud adapter. It is useful when hosted agents or other machines need a filtered read copy and candidate inbox without operating a central host:
pip install "vault-for-llm[supabase]==0.10.2"
vault remote status --project-dir ~/Vaults/my-project
python -m scripts.sync_to_supabase --db ~/Vaults/my-project/vault.db --document-map --health
Gateway / Remote Server is the self-host path, also called the Trusted Local Central Memory Host pattern. Use it when many agents can reach one trusted central memory host:
# Set VAULT_GATEWAY_TOKEN from your shell or secret manager before serving.
vault remote-server health --project-dir ~/Vaults/my-project --json
vault remote-server openapi --project-dir ~/Vaults/my-project --json
vault remote-server serve --project-dir ~/Vaults/my-project --host 0.0.0.0
Remote contributions should enter as review candidates. This is centralized sharing, not offline multi-master sync. Vault Cloud is the future managed backend for the same Vault Governance Contract; it should not change the memory semantics that local, self-hosted, and Supabase deployments use.
The trust model is:
- same machine: several local agents can share one Vault project through local CLI, MCP, Obsidian, or Gateway surfaces;
- different machines or hosted agents: use anon/scoped credentials to read approved memory and submit candidates;
- trusted sync host: the only place for service-role credentials, candidate pull, review/promotion, Dream, archive, forgetting, reports, approved read-copy sync, and derived vector-index updates.
For multi-agent or multi-device memory, use the Central Memory Station surface:
vault start
vault daily-loop run --write-report --json
vault daily-loop report --refresh --write-report --json
vault memory-sync run-once --push-read-copy --push-central-store --pull-candidates --dry-run --json
vault memory-sync run-once --central-backend self-host --pull-candidates --dry-run --json
vault memory-sync migrate-candidates --direction self-host-to-supabase --json
vault memory-sync export-snapshots --bundle ./vault-snapshots.json --json
vault memory-sync verify-snapshots --bundle ./vault-snapshots.json --json
vault memory-sync import-snapshots --bundle ./vault-snapshots.json --json
vault memory-review inbox --json
vault memory-lifecycle status --json
vault ops security --json
vault daily-loop run is the daily orchestrator: it refreshes sync freshness,
automation cycle handoff, inbox, review cards, learning health, and the human
daily report in one report-first pass. Use
vault daily-loop report --refresh --write-report when you only need to rebuild
the latest daily-loop report from read-only status surfaces without capture or
new candidate writes. The detailed remote, sync, dream, usage, and
automation commands stay available as advanced tools. The central surface is
the smaller operator entry for Supabase or self-hosted deployments.
Docs:
- Deployment modes
- Supabase setup
- Supabase read policy
- Gateway security foundation
- Central Memory Station decision
- Single-host sharing and multi-host governed sync
- Central Memory Station plan
- Self-host Central Memory Host spec
- Memory foundation architecture review
Memory Migration
Import memory from other tools as candidates, not as trusted active memory:
vault import memory --source ~/Downloads/chatbox-export.json --format auto --dry-run
vault import memory --source ~/Downloads/chatbox-export.json --write-candidates --only summaries,decisions,preferences
Imported items pass the same privacy, duplicate, metadata, and quality gates.
For backend moves or a self-host bootstrap, reviewed snapshot bundles can move approved memory evidence without direct active-memory writes:
vault memory-sync export-snapshots --bundle ./vault-snapshots.json --json
vault memory-sync export-snapshots --bundle ./vault-snapshots.json --include-content --json
vault memory-sync verify-snapshots --bundle ./vault-snapshots.json --require-content --json
vault memory-sync import-snapshots --bundle ./vault-snapshots.json --apply --json
By default snapshot bundles exclude raw memory content. --include-content
should be used only over a trusted, encrypted transfer path. The verify command
checks the bundle manifest, snapshot digest, content hashes, and missing-content
state without writing memory. Import still writes local memory_candidates for
review; it does not write active knowledge or promote memory.
Retrieval Quality
Vault includes Search QA so retrieval can be measured instead of trusted by intuition alone.
vault search-qa run \
--qa-file benchmarks/search_qa/basic.en.json \
--mode keyword \
--output /tmp/vault-searchqa.json
Current public claims should be read as retrieval evidence, not final answer quality:
- project onboarding proof runs found source-backed memory across 28/28 tasks
- LoCoMo / LongMemEval external-memory probes report retrieval-only source-hit metrics under fixed top-k and source-id matching
- official answerer/judge scores are separate and require model-provider runs
More detail:
Maturity
Vault Agent Memory is pre-1.0. The core local path is intentionally conservative. Advanced remote and automation paths are powerful, but should be enabled deliberately.
Known limitations for the Public Beta / Developer Preview:
- Supabase, Gateway, and central semantic search are optional advanced paths, not required for the local Vault source of truth.
- Remote agents should not receive service-role keys; they should use anon or scoped credentials for approved reads and candidate submission.
- Remote semantic search is disabled by default. If enabled, the default query embedding provider is OpenAI, so search query text is sent to OpenAI unless the operator configures a local or otherwise trusted embedding provider.
- Central vectors index reviewed safe summaries and previews only; they do not make Supabase the active-memory authority.
- Vault Cloud is a future managed backend for the same governance contract, not a replacement for local/self-hosted/Supabase deployment modes.
- LoCoMo / LongMemEval results are retrieval-only source-hit probes unless an explicitly comparable final-QA answerer and judge run is published.
- Automation is report-first and candidate-first. Governed-auto can close routine low-risk metadata/dedup review noise, but strategic, sensitive, freshness, convergence, and consolidation work stays visible for review.
Documentation Map
- Core concepts in plain language
- Deployment modes
- Agent install runbook
- CLI reference
- Agent integrations
- Local GUI console
- Memory governance
- Automation
- MCP tool reference
- MCP workflow
- LLM integration
- OKF integration
- Vision notes
Development
Common Python path:
git clone https://github.com/zycaskevin/Vault-Agent-Memory.git
cd Vault-Agent-Memory
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev,mcp]"
pytest -q
Reproducible Agent/developer environment:
git clone https://github.com/zycaskevin/Vault-Agent-Memory.git
cd Vault-Agent-Memory
uv sync --extra dev --extra mcp
uv run pytest -q
pip install vault-for-llm remains the public user install path. The uv workflow
is for source development, CI smoke checks, and agents that need to rebuild the
same local environment reliably.
Contributing
Vault is ready for small, bounded contributions from Agent-assisted builders. Start with CONTRIBUTING.md, the good first issue ideas, and the Code of Conduct. Please do not include real secrets, private chats, customer records, medical data, or production vault exports in public issues or pull requests.
License
Apache-2.0. See LICENSE.
