HomeProductscaveauMEM
caveauMEM

Persistent memory for any MCP agent.

caveauMEM gives Claude Code, Codex, Cowork, and Copilot a memory that survives the session — storage, retrieval, and maintenance across three tiers, sovereign and scoped, with role-based security and an encrypted secret vault.

Agentic coding has no memory.

Every session starts from zero — and the workarounds leak, sprawl, and go ungoverned.

01

Agents forget everything

Every new session starts cold. Hard-won context about your repos, decisions, and preferences evaporates the moment the window closes.

02

Context is re-pasted forever

Teams burn time and tokens re-explaining architecture, deploy steps, and house rules into prompts that vanish on the next run.

03

Secrets leak into prompts

API keys and tokens get pasted inline to "just make it work" — ending up in logs, history, and provider context with no governance.

04

No governance over shared memory

Once a team shares "company" memory, there is no scoping, no review, and no way to keep confidential work out of the wrong hands.

One memory endpoint for every agent you use.

caveauMEM speaks MCP, so it plugs into Claude Code, Codex, Cowork, and Copilot through a single endpoint — your memory travels with you, host to host.

One memory_bootstrap call at the start of a session loads, in a single shot:

  • The repo’s rules, architecture, and deploy notes
  • Your global personal working preferences
  • The key names (never values) of secrets relevant to the work
  • Active / archive / ignore status for the project

Durable, ephemeral, and episodic — each for what it does best.

caveauMEM keeps long-term knowledge, short-term scratch, and a replayable timeline distinct, so the right memory is recalled at the right moment.

Tier 02

Ephemeral scratch

Fast key/value working memory with a TTL — the agent's short-term notepad for the task at hand. Set it, read it, drop it when the job is done.

  • Key/value set, get, delete
  • Per-key TTL expiry
  • Scoped per client and per user
  • Never pollutes the durable corpus
Tier 03

Episodic timeline

An append-only log of what happened and when — a replayable trail of actions and outcomes the agent can recall to stay coherent across long-running work.

  • Append-only event log
  • Time-scoped recall
  • TTL-bounded retention
  • Plus a to-do / task tier for open work

Shared memory you can actually govern.

Memory is only safe to share when you control who sees what. caveauMEM scopes every note by client, user, and role — and keeps secrets in a vault, never in a prompt.

Tenant

Per-client isolation

Every client is a hard boundary. One tenant's corpora, scratch, and episodic memory are never visible to another.

Personal

Per-user vs shared memory

Separate a user's private notes from the shared "company" memory the whole team draws on — without mixing the two.

Roles

Role → scope matrix

Owner, editor, tech, and viewer roles map to exactly what each person can read, write, and promote.

Confidential

Classified corpora

Mark corpora internal or confidential. Client-work corpora integrate with caveauCRM so memory follows the account.

Vault

Encrypted secret vault

Secrets live in an AES-256-GCM vault, fetched by exact key only — never embedded in notes, never searchable, never logged.

Governed

Propose → promote review

New shared memory enters a review queue. A human promotes it before it becomes part of the team's trusted knowledge.

Give your agents a memory worth keeping.

Start free in the caveauAI portal, or book a call to scope a sovereign, scoped memory deployment for your team.

Technical specifications & tools

Tool inventory

memory_bootstrap
Called once per repo at session start — returns that repo's rules, architecture, and deploy notes plus personal preferences and relevant secret key names.
memory_save
Persist a durable, searchable note to the memory corpus, tagged by repo / client / type / tool.
memory_recall
Semantic search across saved notes, returning ranked results with source citations.
memory_recall_repo
Deterministic, tag-scoped recall of a single repo's notes, filterable by type.
memory_update
Revise an existing note in place, keeping its identity and tags.
memory_archive
Retire a note from active recall without destroying the record.
memory_dedup_scan
Read-only: surface near-duplicate note pairs by cosine similarity for human cleanup.
memory_staleness_scan
Read-only: flag type-tagged notes untouched for N days so curators can refresh them.
memory_secret_set
Store a secret in the encrypted vault (AES-256-GCM at rest).
memory_secret_get
Fetch a secret by exact key for use in a task — never embedded or searchable.
memory_secret_list
List the key names (never values) of secrets relevant to the current scope.
memory_scratch_set / get / delete
Ephemeral key/value working memory with a per-key TTL.
memory_episodic_log / recall
Append to and replay the time-scoped episodic timeline of actions and outcomes.

Capabilities

Hybrid recall Dense embeddings + BM25 lexical, fused with Reciprocal Rank Fusion (k=60) and a private-scope boost.
Graph recall Traverse links between notes to pull in related context, not just the top-k matches.
Context assembler Token-budgeted assembly packs the most relevant memory into the model's window without overflow.
To-do / task tier Track open work items alongside knowledge, so the agent knows what is still outstanding.
Scoping Per-client (tenant) isolation plus per-user vs shared "company" memory.
Role matrix Owner / editor / tech / viewer roles map to read, write, and promote permissions.
Classification Corpora marked internal or confidential; client-work corpora integrate with caveauCRM.
Token scopes MCP tokens carry memory.read / memory.write / secrets scopes — least privilege per host.
Tag taxonomy repo: / client: / type: / tool: tags drive deterministic, filterable recall.
Encryption Secret vault is AES-256-GCM at rest; exact-key fetch only, never embedded or searchable.
AI Chat — Beta Testing, Online Soon