Works with the AI Agents you already use
AI Memory Architect
Design selective, permissioned long-term agent memory with provenance, retrieval, correction, retention, and deletion controls.
$9.99
AI Memory Architect
Example session with this skill installed
We are designing a multi-tenant project assistant. It should remember project decisions and a user’s explicit formatting preferences, but must not retain secrets, sensitive personal attributes, or transient status as durable identity. Users need to correct and delete memories. Propose memory types and schemas, write/consent rules, provenance/confidence/expiry fields, retrieval and conflict handling, tenant isolation, deletion propagation, evaluation metrics, and rollout tests. No existing architecture or policies have been provided, so label assumptions and open decisions.
- Read your context and instructions
- Compiled the ai memory architect
Project Assistant Memory Architecture — Initial Proposal
Summary
Use selective, typed memory with explicit scope, source, time, and validity. Keep user preference, project decision, procedure, and session history separate. Require visible correction and deletion paths. This is a design proposal only: no current storage, identity, consent, or retention settings were inspected.
Memory record model
| Field | Purpose |
|---|---|
| memory_id, tenant_id, subject_id | Stable identity and strict user/project scope; validate on every read and write |
| type, key, value | Preference, project decision, procedure, or episodic reference; avoid free-form dumping |
| source_ref, learned_at, source_type | Provenance and date first observed; never imply an unavailable conversation was retrieved |
| confidence, status | Explicit, confirmed, inferred, disputed, superseded, or deleted; explicit user statements rank above inference |
| valid_from, expires_at, last_confirmed_at | Time validity; preferences and project state can become stale |
| consent_basis, sensitivity_class | Record permission basis; reject secrets and sensitive attributes by default |
| retention_rule, deletion_state, derived_from | Lifecycle control and traceability through summaries, embeddings, caches, and indexes |
Capture and write policy
- Candidate must have expected future utility and durability; ask before storing an inferred preference or ambiguous project fact.
- Store explicit formatting preferences only within the user scope; store project decisions only in the project scope with date, source, decision owner, and status.
- Do not persist credentials, secret keys, sensitive personal traits, transient moods/status, or third-party private details. Treat user-provided sensitive content according to approved policy and minimize retention.
- Show the user what will be remembered, why, and for how long where the product supports it. Offer correction, forget, pause, and review controls.
- Validate schema, tenant/subject identity, deduplication, consent, and retention before commit. Confirm success before saying it was saved.
Retrieval and conflict handling
Filter by authorized tenant, user, project, purpose, status, and validity before semantic ranking. Prefer current explicit corrections over stale or inferred records. Apply relevance plus freshness and confidence; do not treat similarity alone as truth. When two active project decisions conflict, preserve both sources, mark the conflict, ask the owner, and avoid silently merging. Include source and timestamp in assembled context; retrieve the minimum necessary records.
Correction, expiration, and deletion
Corrections create a versioned record and mark the prior entry superseded with provenance. Expire or decay mutable preferences and project state under owner-approved periods; do not refresh simply because a record was retrieved. On deletion, remove or tombstone the source record and propagate to vector indexes, caches, summaries, replicas, and derived memories per policy. Verify with deletion-job evidence and a scoped re-query; do not claim complete deletion until all required stores confirm.
Evaluation and rollout
Test cross-tenant and cross-user access, purpose mismatch, stale preference, conflicting project decisions, unsupported inferred memory, secret exclusion, correction propagation, deletion across derived stores, and retrieval relevance. Measure task success and repeat-question reduction alongside stale-memory rate, unsupported-memory assertions, leakage rate, correction success, deletion completion, latency, and token cost. Use synthetic data first; pilot with explicit opt-in, inspect outcomes with owners, and retain rollback.
Open decisions
Confirm consent model, permitted memory types, retention periods, identity source, tenant/project hierarchy, storage providers, deletion SLA, legal/privacy obligations, high-risk data exclusions, user controls, and owner for exceptions. Until decided, default to session-only context and do not persist.
Connects securely to your tools. The creator never sees your data.
About this skill
AI Memory Architect
Design, audit, and improve long-term context systems for AI assistants and agents. Retain the smallest set of trustworthy, permissioned information that materially improves future assistance; do not turn memory into exhaustive conversation storage.
What it helps design
- Memory capture policies and candidate selection for preferences, project state, decisions, procedures, and history.
- Structured and semantic storage, identity/tenant scope, source provenance, confidence, timestamp, and validity.
- Retrieval, ranking, temporal reasoning, context assembly, deduplication, contradiction handling, and consolidation.
- Correction, update, decay, archival, forgetting, retention, and deletion propagation across indexes, caches, summaries, and derived records.
- Consent, sensitive-data exclusion, access controls, utility and quality evaluation, latency, and token-cost monitoring.
Design principles
Utility must justify persistence. Explicit instructions and current corrections outrank inference and stale records. Separate memory types and source facts; preserve when, where, and how a memory was learned. A retrieved memory is evidence of past context, not automatically current truth. Verify mutable facts when current accuracy matters. Enforce user and tenant boundaries, ask consent where required, and never store secrets merely because they appeared. Claim save, update, or deletion only when the system confirms it.
Use it well
Describe users, agent purpose, sessions, memory goals, current context/storage, identity and tenant model, memory types, capture sources, consent, sensitivity, retention, deletion, retrieval, embeddings, summaries, project state, evaluation data, and scale constraints. If facts are missing, state conservative assumptions and list what owners must decide.
Boundaries
This skill designs architecture and evaluation plans. It does not assume access to past conversations, invent user facts or permissions, or claim memories were saved or deleted without system evidence. Production retention, privacy, legal, and security decisions require qualified owners.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for LLM Ops.
Architecture recommendations depend on accurate information about the memory sources, identity model, consent, storage, and retention rules. Retrieval quality and deletion completeness cannot be guaranteed without testing every connected store, index, cache, replica, and derived summary. Privacy and legal requirements need review by the responsible owners.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 1 month ago
Permissions
Use file access only when the user authorizes it and limits the requested scope. Do not read or persist personal memory by default; do not claim production memory changes, privacy compliance, deletion, or connector access without verified system evidence.