ASM-CM: Compact Persistent Memory for AI Agents Without Keeping the Full History Active

Hi everyone,

I’m Felipe Maya Muniz, founder of AletheionAGI and creator of ASM — Aletheion State Models.

I’m developing ASM-CM, the Aletheion Compact Memory Model, an experimental memory architecture for persistent AI agents.

The central research question is:

Can an agent’s history keep growing without its active computational memory growing at the same rate?

ASM-CM does not attempt to replace the language model. Instead, it separates memory from language:

Events and observations
        ↓
ASM-CM
compact state, associative memory and retrieval
        ↓
Minimal relevant context
        ↓
Local or remote LLM
language and interpretation

The LLM produces the response. ASM-CM preserves and retrieves associations.

A deterministic application or world model remains responsible for canonical truth, permissions and actions.

What Is It Intended For?

Possible applications include:

• persistent NPCs and game worlds;
• long-running AI agents;
• customer-service continuity across sessions;
• private organizational assistants;
• local memory for remote LLMs;
• robotics and long-running processes;
• continuity controllers for long-form audio, video and interactive media;
• per-user or per-agent isolated memory;
• systems that need to change the LLM without losing the agent’s memory.

ASM-CM is not a database replacement. A practical application may still use databases, event logs, RAG and knowledge graphs.

Its intended role is different:

Maintain a compact active state and recover only the associations relevant to the current interaction.

The Diamond Promise Experiment

To make the behavior understandable, we created a small persistent-world experiment called The Diamond Promise.

Day 1: Deposit

Aldric gives a diamond to Elana and says he will return in three days.

The deterministic world records:

• Aldric remains the legal owner;
• Elana becomes the current custodian;
• Elana remembers the deposit and the promise.

Day 2: Robbery

Bram threatens Elana and steals the diamond.

The world updates the canonical state:

• Aldric remains the legal owner;
• Bram becomes the canonical holder;
• Elana no longer possesses the diamond.

Elana remembers the robbery because she witnessed it.

Aldric does not automatically acquire that knowledge because he was absent.

Interference

The system then processes thousands of unrelated events per character.

These distractors test whether the relevant associations survive interference instead of being displaced by everything that happens afterward.

The memory is also persisted to a snapshot, the process is terminated and another process restores it.

Day 3: Retrieval

Aldric returns and asks Elana:

Where is my diamond?

ASM-CM retrieves two relevant memories from Elana’s state:

Aldric deposited the diamond.
Bram threatened Elana and stole it.

Only the selected memories are sent to a local Qwen3 8B model, which produces Elana’s response.

The LLM cannot change ownership, inventory or past events. Its output is evaluated against both the retrieved memories and the deterministic world state.

The final checks include:

• deposit retrieved;
• robbery retrieved;
• both memories cited by the LLM;
• Aldric’s private knowledge remained isolated;
• the no-memory control remained empty;
• legal ownership was preserved;
• canonical holder was preserved;
• memory survived process restart;
• retrieval survived thousands of distractors.

The current matrix passed 15 out of 15 cases, including scenarios with up to 10,000 distractors per character.

The hardest case was subsequently evaluated with three independently trained checkpoints.

Long-Streaming Results

In the frozen 32K streaming protocol, ASM-CM achieved:

• 100% MQAR associative-retrieval accuracy;
• approval across three training seeds;
• approximately 140 KiB of retained state per stream;
• approximately 363.66 MiB peak VRAM for the evaluated ASM-CM component;
• stable retained-state size as sequence length increased.

These numbers apply only to the evaluated memory component and protocol.

A complete system—including an LLM, database, interface and application services—requires additional memory.

The result does not demonstrate infinite memory. A fixed-size state has finite capacity and may experience interference or forgetting.

A more accurate description is:

The history can grow without the active neural state growing proportionally with it.

Endurance Protocol

We are also testing real elapsed-time persistence:

• 1-hour protocol: passed;
• 6-hour protocol: passed;
• 24-hour protocol: currently being finalized;
• 72-hour protocol: not yet demonstrated.

The endurance test includes:

• sustained distractor processing;
• periodic atomic snapshots;
• termination of the original process;
• restoration under a different process ID;
• snapshot-hash verification;
• final associative-memory probe;
• retained-state-size gate.

Results will be reported according to the frozen criteria.

The 24-hour result should not be considered approved until the run and final probe have completed.

Where Could the Savings Come From?

A conventional agent may repeatedly send a large conversation history to an external LLM.

A compact-memory architecture attempts to send only the relevant retrieved context.

A simplified estimate is:

monthly input-token savings
=
requests per month
×
(full-history tokens − selected-context tokens)

For example, consider a system with:

• 10,000 persistent agents;
• 100 interactions per agent per day;
• 10,000 historical tokens normally resent per interaction;
• 1,000 selected tokens sent after memory retrieval.

The difference would be:

10,000 × 100 × 30 × (10,000 − 1,000)
= 270 billion fewer input tokens per month

This is a scenario, not a measured customer result.

Actual financial savings depend on:

• provider and model pricing;
• prompt caching;
• average context length;
• retrieval accuracy;
• local infrastructure;
• request frequency;
• how much historical context the current system already avoids.

ASM-CM would not eliminate all LLM or infrastructure costs.

The relevant commercial question is whether the cost of operating the memory layer is lower than the context, latency and privacy costs it avoids.

Comparison With Transformers and RAG

A Transformer can answer the Diamond Promise correctly if it receives the relevant history.

A RAG system may also retrieve the deposit and robbery from an external store.

Therefore, the claim is not that only ASM-CM can remember.

The differentiating hypothesis is:

• bounded active neural state;
• associative retrieval;
• persistence across processes;
• isolated memory per character or agent;
• independence from the language model;
• minimal context disclosure to local or remote LLMs.

RAG asks approximately:

Which stored documents are similar to this query?

ASM-CM investigates:

Which associations should this agent recover in its current state?

They can be complementary.

A production system could use ASM-CM for active continuity and RAG or databases for durable records and open semantic search.

Current Limitations

ASM-CM has not demonstrated:

• universal superiority over Transformers or RAG;
• general narrative understanding;
• unlimited storage capacity;
• automatic data security;
• open-domain semantic retrieval at production scale;
• superiority in general language modeling.

In our paired language benchmark, the Transformer baseline still achieved better cross-entropy and training throughput.

The demonstrated property is narrower:

Compact, persistent and recoverable associative memory under the evaluated protocols, capable of supplying selected context to an LLM without keeping the complete past active.

Open Source and Collaboration

The source code, tests, reports and reproducible protocols are publicly available under AGPL-3.0-only:

• GitHub: https://github.com/gnai-creator/ASM
• Website and evidence: https://www.aletheionagi.com
• Technical results: https://www.aletheionagi.com/evidence
• ASM-CM overview: https://www.aletheionagi.com/models/asm-cm

Commercial use is permitted under the AGPL-3.0-only when its obligations are satisfied.

Separate commercial licensing is available for proprietary integration, closed-source deployment, alternative terms and enterprise support.

AletheionAGI is open to conversations involving:

• reproducible independent evaluation;
• paid pilots;
• integration with games or agent platforms;
• commercial licensing;
• field-of-use exclusivity;
• joint development;
• research partnerships;
• investment;
• acquisition of the technology or intellectual property.

A good initial pilot would use a small number of agents, a frozen no-memory or RAG baseline and measurable criteria for:

• retrieval;
• false memories;
• latency;
• retained state;
• privacy;
• cost.

If your project has persistent agents that repeatedly carry large histories, I would be interested in understanding the workload and comparing ASM-CM with your current memory architecture.

Contact: [e-mail]