AI agent memory
Agent memory is not a longer prompt.
Every team hits the same wall. The AI does good work in a session and forgets it by the next one, so the same context gets pasted in again, slightly differently each time. Stuffing more into the prompt delays the wall; it does not remove it.
Available now
Composer, public Guard scans, workspace records, policy, activity metadata and scoped keys.
Preview
Skill and Agent execution, Memory retrieval, schedules, publishing, billing and unified execution.
Context
Three different things get called memory
They have different lifecycles, different owners and different risks. A system that treats them as one blob is why retrieval returns confident nonsense.
Knowledge — what you gave it. Documents, websites, repositories, notes. You control it; it changes when you change it.
Memory — what it learned. Decisions, project state, prior outcomes, useful summaries. It accumulates through use and needs pruning.
Rules — what always applies. Brand tone, coding standards, terminology, prohibited actions. Short, stable, and should never be crowded out by retrieval.
Evaluation
What to check in a memory layer
Retrieval quality is easy to demo and hard to trust. These are the properties that make it usable at work.
Can it tell you where an answer came from?
Source references with an anchor back into the document. Without citation you cannot check the answer, and you will not be allowed to use it for anything that matters.
Is access granted per agent, not per workspace?
Tenant isolation works today. Per-collection and per-Agent grants are part of the planned access model and are not yet enforced.
Does a failed index fail loudly?
Source records and their status remain visible today. Indexing and retry behavior are still planned because no retrieval index is built.
Are rules separated from knowledge?
Standing instructions must always apply. If they compete with document chunks for retrieval space, they will lose exactly when they matter.
Can you delete and export?
Individual collection and source records can be removed. Full export, backup policy and a complete Memory retention workflow are not implemented.
Which models can see it?
The tenant's AIHK/AICA region lock applies today. Per-collection model-access rules remain part of the planned access model.
GenCMS
How GenCMS Memory is designed
Knowledge + Memory + Rules = Context. Collection and source records persist today; indexing, retrieval, citations, grants and run-time Memory access are not implemented.
Sources
Store supplied text and source records for URLs, files, repositories and structured data. Remote content and files are not fetched, parsed or indexed.
Source records →Collections
Create named, tenant-scoped knowledge, memory or rules collections and persist their source records.
See Memory →Citation
Retrieved source references and anchors remain part of the planned retrieval layer; no citation result is produced today.
Planned retrieval →Access control
Workspace tenancy is enforced today. Per-collection grants, retrieval audit, export and execution-time Memory permissions are not implemented.
Planned access model →Questions
Asked before you commit.
Is this RAG?
Not yet. The current workspace stores collections, supplied text and source records, but it does not chunk, embed, index or retrieve them. The proposed retrieval design separates supplied knowledge, retained context and standing rules.
What file types can I add?
You can store supplied text and source records for URLs, files, repositories and structured data. GenCMS does not currently upload, fetch, parse or index the underlying file or remote content.
Can an agent write to memory?
Not today. Agent execution, collection grants and memory.write are product-preview concepts.
Where is the data stored?
Memory collection and source records are stored in the tenant-scoped workspace data store. There is no retrieval index, and production storage, backup, export and retention arrangements must be agreed for a deployment.
Explore the Memory design.
Store collection and source records today, then review the indexing, citation and access layers still on the roadmap.
GenCMS is a product of IGEARS TECHNOLOGY LIMITED. Guard results describe observed signals, not guarantees.