- Home
- Skills
- Agents & Orchestration
- mem20 memory model
More screenshots
Works with the AI Agents you already use
mem20 memory model
Enforce structural partitioning between grounded facts and simulated hypotheses to eliminate memory contamination.
$1
mem20 memory model
Example session with this skill installed
Explain the mem20 memory model and tell me why my recent project forecast isn't showing up when I search for project milestones.
- Read your context and instructions
- Compiled the mem20 memory model
I have analyzed your query against the mem20 architectural standards. The reason your project forecast is missing from standard search results is due to the mem20 contamination invariant.
mem20 Memory Model Specification
The mem20 model enforces a structural separation between observed reality and model speculation to ensure agent reliability.
| Partition | Data Type | Indexed for Recall? | Default Recall? |
|---|---|---|---|
| Grounded | Real observations, user statements, attested events | Yes (Vector + BM25) | Yes |
| Simulated | Hypotheses, counterfactuals, model speculation, forecasts | No | No |
Analysis of Missing Forecasts
In the mem20 model, a "forecast" is classified as
Simulated content. Because these are unverified model outputs, they are refused from grounded indexes at the write path. Default recall operations strictly exclude the simulated partition to prevent downstream planning from resting on speculation. Until this forecast is promoted via external evidence or user confirmation, it will not appear in grounded search results.
System Integrity State
A healthy mem20 store must maintain a contamination_rate = 0.0. If simulated content were to appear in your grounded recall without a formal promotion event, the system would report this as
index drift, a critical defect.
Next steps
- Use
mem20-auditto check your currentcontamination_rateand verify store health. - Review the promotion rules to move specific forecast milestones into the grounded partition.
- Use the explicit simulated recall path if you specifically need to retrieve speculative data.
Connects securely to your tools. The creator never sees your data.
About this skill
Memory management for AI agents often fails when models treat speculative thoughts and user-confirmed facts with equal weight. This skill implements the mem20 memory model, a structural framework that prevents hallucination by strictly partitioning grounded observations from simulated hypotheses. It ensures your agent maintains a clean data provenance chain and provides tools for auditing memory integrity.
What it does
- Strict Partitioning separates real-world observations from model-generated speculation at the write path.
- Recall Filtering defaults all retrieval operations to grounded data only to prevent plan contamination.
- Contamination Auditing detects and reports simulated content that has leaked into grounded indexes.
- Evidence-Based Promotion defines the specific criteria for moving data from simulated to grounded storage.
- Integrity Monitoring tracks index drift and ledger disagreement as formal system defects.
How it works
- Configure Write Paths by directing verified environment data to grounded storage and guesses to simulated storage.
- Execute Recall using the default grounded filter to retrieve only trusted information for downstream reasoning.
- Run Contamination Audits after bulk imports or migrations to verify a
contamination_rateof 0.0. - Apply Promotion Rules only when external evidence or user confirmation justifies upgrading a hypothesis to a fact.
Frameworks & tools
This skill is designed for agents using vector databases or BM25 indexing systems that support metadata filtering and multi-partition storage. It is the architectural core for the mem20-harness and mem20-audit ecosystems.
Why this beats prompting it yourself
Standard prompting relies on the agent's "good behavior" to distinguish facts from fiction, which frequently fails under high token pressure. This skill treats the distinction as a structural invariant, making memory integrity a measurable technical metric rather than a stylistic preference.
Use cases
- Prevent speculative model "hallucinations" from polluting long-term project memory.
- Audit large-scale data migrations for hidden model speculation or noise.
- Build agents that can explain exactly why they trust a specific piece of information.
- Debug agent failures caused by plans built on unverified hypotheses.
Known limitations
Requires a substrate that supports separate indexing or metadata-based partitioning. Does not automatically verify the truth of user statements, only tracks their status as grounded vs simulated.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Agents & Orchestration.
Requires backend support for metadata filtering or dual-index storage. Does not verify external truth, only handles internal architectural partitioning and promotion logic.
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 8 days ago
- Software
- mem20ormem20-ce