Works with the AI Agents you already use
Agent Memory Architect — Design Useful AI Memory
Design persistent context that is useful, selective, privacy-aware, traceable, correctable, and safe to retrieve across AI agent sessions.
$9.99
Agent Memory Architect — Design Useful AI Memory
Example session with this skill installed
Design a persistent memory policy for a project-management copilot used by multiple client teams. It should remember stable team preferences and project decisions, but not turn every chat into memory. Teams must never see one another’s data. Decisions can be superseded; client contacts may request deletion. We have a relational database and application authentication, but no vector store selected. Specify memory classes, candidate-write rules, schema fields, retrieval triggers, conflict/supersession, expiry and deletion, tenant isolation, user review controls, evaluation tests, and rollout questions. Separate supplied facts from assumptions and do not claim any specific platform memory feature.
- Read your context and instructions
- Compiled the agent memory architect
Persistent memory blueprint: multi-client project copilot
Executive summary
Treat memory as a curated, tenant-scoped record of durable context—not a copy of chat history. Start with explicit write rules, provenance, user review, and deletion propagation. Keep operational status in the relational system of record; retrieve only the smallest relevant memory set for a task.
Inputs and assumptions
Supplied: project-management copilot; multiple client teams; remember stable preferences and decisions; avoid saving every chat; strict team isolation; decisions may be superseded; client contacts may request deletion; relational database and application authentication exist; no vector store selected.
Assumptions to validate: the app can identify tenant and project on every request; a user or authorized administrator can review and request correction/deletion; legal retention requirements and backup deletion behavior are not yet specified.
Memory classes and write policy
| Class | Example | Candidate write rule | Default lifetime |
|---|---|---|---|
| Team preference | Preferred status format | Explicitly confirmed or repeated and user-approved; scoped to one tenant/team | Review yearly or when corrected |
| Project decision | Approved launch date | Record only with source, author, timestamp, scope, and status; newer approved decision supersedes | Keep decision history per retention policy |
| Workflow state | Current task status | Do not duplicate if authoritative in the project database; query that system | System-of-record lifecycle |
| Interaction episode | One-off discussion | Do not persist by default; extract only durable approved facts | Short-lived or none |
| External knowledge | Policy or product fact | Keep source-grounded with version and effective date; do not treat retrieval as truth | Expire or refresh by source cadence |
For each candidate, evaluate future value, confidence, durability, sensitivity, user intent, scope, and cost of being wrong. Reject transient, uncertain, sensitive, or low-value facts unless explicit policy and consent permit storage.
Record model
Store: memory ID; tenant ID; project ID; class; normalized content; source reference; source type; author/actor; created and observed times; confidence/evidence status; sensitivity label; consent or write-policy basis; effective-from/to; status (active, superseded, expired, deletion-pending); supersedes ID; review date; and audit events. Keep content minimal and never use a tenant-global index without tenant filtering enforced server-side.
Retrieval and conflict rules
Trigger retrieval only when the task needs continuity (for example, planning a project using prior approved decisions). Filter by authenticated tenant and project before ranking. Rank active memories by scope, recency, authority, relevance, and confidence. Return provenance with each item. A current explicit correction outranks an older memory; preserve a supersession link when policy requires history. If two active decisions conflict, surface both with dates and ask which governs rather than silently merging them.
Expiry, review, and deletion
Assign class-specific review dates. Expire preferences that are no longer confirmed and time-bound facts at their end date. For a deletion request, identify derived copies, indexes, caches, and backups covered by policy; mark deletion state and record completion evidence. Define whether legal holds or backup retention impose exceptions before promising deletion timelines. Provide users a way to inspect, correct, and remove eligible memories.
Security and evaluation
Authorize every read and write against tenant/project membership; test cross-tenant retrieval and guessed-ID access (expected zero leakage). Validate write precision, duplicate rate, correction propagation, stale-memory rate, retrieval relevance, deletion completion, latency, and task success with and without memory. Include adversarial tests for prompt-injected content, untrusted third-party claims, mistaken identity, and conflicting instructions.
Rollout
- Inventory tasks, data classes, retention obligations, identity and tenant model.
- Pilot explicit user-confirmed preferences and project decisions in one test tenant.
- Add review, correction, expiry, and deletion workflows before broad ingestion.
- Benchmark relational filtering and keyword retrieval on representative tasks before selecting a vector store.
- Roll out gradually with tenant-isolation gates and an auditable rollback path.
Open questions
Who may approve memory writes? Are client contacts users or recorded subjects with deletion rights? What retention and legal-hold rules apply? How are backups purged? Which source is authoritative for project state? What languages, request volume, and latency targets are expected?
Connects securely to your tools. The creator never sees your data.
About this skill
Design memory that earns its place
Agent Memory Architect helps AI builders decide what an agent should remember, why it matters, where it belongs, when it should be retrieved, how it is corrected, and when it should expire or be deleted. It distinguishes conversation context, durable preferences, project memory, client history, decision records, workflow state, episodic history, semantic knowledge, source-grounded knowledge, operational databases, temporary working memory, and derived summaries.
What it designs
- Persistence value and privacy tests for candidate facts
- Memory classes, scope, ownership, tenant boundaries, and provenance
- Write policies, consent, confidence, source and timestamp capture
- Storage schemas, retrieval triggers, ranking, and minimal context assembly
- Deduplication, consolidation, corrections, conflicts, and supersession
- Expiration, review, user controls, deletion, retention, and portability
- Poisoning defenses, false-memory prevention, access controls, and auditability
- Evaluation plans that test whether memory improves future task outcomes
Design principle
Memory is not a transcript archive. Persist only information likely to improve future work enough to justify its privacy, correctness, and maintenance cost. Keep changing operational facts in systems of record and external knowledge source-grounded. Never claim a platform can save, retrieve, edit, or delete memory unless that capability is actually available. Never fabricate remembered facts.
Best results
Share the agent's purpose, users, cross-session tasks, current memory behavior, available storage and tools, authorization model, data sensitivity, retention and deletion requirements, expected scale, and representative success cases. Remove secrets and unnecessary personal data.
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.
Memory designs depend on actual storage, authentication, consent, retention, and deletion capabilities. A blueprint cannot guarantee isolation, successful deletion, or retrieval quality without implementation and testing. Legal obligations, backup handling, user roles, and data residency require confirmation from the system owner. Recommendations are not a substitute for platform-specific security review.
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
No external services are required by default. Read only architecture briefs or policies explicitly supplied or authorized by the user. Do not persist, query, modify, or delete live memory or personal data unless the user explicitly authorizes the specific action and the target environment supports it.