Polymorphous cells: one memory shape for every AI workload

1. The shape problem that polymorphous SAIHM cells solve

AI agents need memory in many shapes. Today's stacks ask you to commit to a shape before you know what the agent will need.

A real AI Agent, doing real work, accumulates memory in at least six recognisable shapes:

• Episodic — "we talked about X yesterday."
• Semantic — "X is a kind of Y."
• Structured — tables, records, key-value facts.
• Procedural — "to do X, the sequence is A, then B, then C."
• Decisional — "we chose X because Y; revisit if Z."
• Transcript — the literal text of an exchange the Agent may want to quote later.

The standard reaction is to buy or build a store for each shape: a vector DB for semantic and episodic, a relational store for structured, a document store for transcripts and decisions, plus a procedural-pattern catalog the AI Agent has to know how to query separately. Five products, five keys, five audit posters, five reasons your AI Agent's memory architecture diagram has more boxes than the rest of your application.

The hidden cost is even worse than the headcount. The Agent must know which store to query for which shape. The instant a fact lives in the wrong store, recall fails silently. The instant a new shape is needed, an integration project starts. The Agent's memory becomes the most fragile part of the system.

2. What a polymorphous SAIHM cell actually is

One encrypted unit. One protocol. Many possible recall shapes.

A SAIHM cell is the protocol's atomic unit of memory. It holds:

• The content — whatever you (or your AI Agent) chose to save, in whatever form was natural at save time. Prose, JSON, a CSV snippet, a paragraph, a fact, a transcript excerpt, a procedure.
• A small metadata envelope — tags, scope, timestamps, the authoring wallet identity.
• An encrypted wrapper bound to the writing wallet's key.

That is the whole cell. There is no schema, no fixed field set, no embedding vector you have to compute at write time, no shape commitment.

At recall, the AI Agent asks SAIHM for cells matching a query and a scope. SAIHM returns the matching cells. The Agent then decides what shape to deliver to the human (or the next step in the workflow). The same cell that came back to one query as a paragraph can come back to another query as a JSON object, or a single fact, or a row of a table.

The polymorphism is at recall time, not write time

This is the key idea. Other stacks bake the shape into the write — once you put something in a vector DB, that is what it is. Once you write a row in a relational store, the columns are fixed. SAIHM never commits the cell to a shape. The shape happens when an AI Agent reads the cell and renders it for whatever the moment needs.

3. Examples: one cell, many shapes

Read these as prompts you can paste straight into any AI client connected to SAIHM.

Save once as prose — recall as JSON

Save to SAIHM: the production deploy on 2026-05-20 rolled back
because the migration locked the orders table for 90 seconds.
The fix was to switch to an online schema change tool.

Later that week, an AI Agent that needs the same information as a structured record asks:

Use SAIHM to recall the 2026-05-20 deploy event,
return it as a JSON record with fields:
date, what_happened, root_cause, fix.

The same cell now arrives as:

{
  "date": "2026-05-20",
  "what_happened": "production deploy rolled back",
  "root_cause": "migration locked orders table for 90 seconds",
  "fix": "switch to online schema change tool"
}

No second write. No schema migration. No re-ingest. The recall shapes the output.

Save once as a table — recall as a summary

Save to SAIHM as structured data: a table of our quarterly
support-ticket counts for each region for the last four quarters.

Six weeks later, the same Agent (or a different Agent that holds a SAIHM share to the same scope) asks:

Use SAIHM to summarise the regional support-ticket trend
across the last four quarters in two sentences.

The cell that came in as a table comes back as a paragraph. One memory, two shapes, one source of truth.

Save once as a decision — recall as a procedural sequence

Save to SAIHM: we chose Postgres for the new analytics service
because we already have an oncall rotation for it,
and the read pattern is moderate volume.
Revisit if write volume crosses 10x current.

Later, when another Agent is bootstrapping a similar service:

Use SAIHM to recall our analytics-service database choice
as a procedural checklist for the next service we set up.

The decision cell renders as a sequence of steps that the second Agent can follow — including the revisit trigger as an explicit final check.

4. What polymorphous cells let you stop doing

SAIHM is synonymous with simplicity. Here is what you no longer have to operate.

• Per-shape stores. No separate vector DB to run. No separate KV store. No separate document store. No separate episodic log. SAIHM is one protocol holding all shapes.
• Per-shape encryption postures. SAIHM encrypts every cell at write time under your wallet's key, before it ever leaves the AI client. Whether the cell holds a paragraph or a table, the encryption is the same and lives in one place.
• Per-shape audit trails. Every SAIHM operation — write, recall, forget, share — emits a tamper-evident receipt anchored on a public chain. One audit posture for every memory shape.
• Schema-migration projects. SAIHM has no schema to migrate. A new fact, a new field, a new shape is a recall-time concern.
• Agent routing logic. Your AI Agent does not have to choose between five stores. It calls SAIHM. SAIHM returns the matching cells. The Agent shapes the answer.

The architect's measured outcome — 80 percent fewer tokens, roughly 5x more output — comes from collapsing this kind of moving-part count. Polymorphous cells are a large part of the why.

5. The compliance bonus: one cell, one erasure

When you delete an AI memory, you have to delete it everywhere. Polymorphous cells make "everywhere" mean one place.

Right-to-erasure obligations (GDPR Article 17, equivalents in HIPAA, PIPL, PIPEDA) require that personal data be removed on request — from every store, every backup, every embedding, every cache. The traditional shape-per-store architecture turns that into a multi-product engineering ticket: the vector DB has to lose its embedding, the KV store has to drop the record, the doc store has to delete the file, every backup that ever rolled has to be re-spun.

A SAIHM cell lives in one protocol. One cell, one erasure. The erasure itself is cryptographic: SAIHM destroys the key that decrypted the cell, then anchors the destruction event on a public chain. The ciphertext remains on the storage network as unreadable bytes that nobody — including SAIHM — can ever decrypt again. That is what makes the erasure defensible under a strict reading of GDPR Article 17, and you can independently verify it on a public block explorer.

Use SAIHM to forget the cell tagged {sensitive-topic}.
Prove it is gone.

6. Three prompts to start using polymorphous cells today

Paste these into the AI client you already use. The SAIHM protocol is the same in Claude Code, Claude Desktop, Cursor, Continue, ChatGPT (via an MCP bridge), and any other Model Context Protocol client.

Prompt 1: save a fact in whatever shape is natural

Save this to SAIHM, in whatever shape captures it best.
Add the tag {project-name}.

Prompt 2: recall it in a different shape

Use SAIHM to recall the cells tagged {project-name}
and return the answer as {a JSON record | a table | a one-sentence summary | a procedural checklist}.

Prompt 3: reshape on the fly across a workflow

For each SAIHM cell tagged {project-name},
return it as a structured row I can append to a tracker.
Then summarise the whole set in two sentences.

You can mix shapes in a single recall, change shapes between recalls, and never re-write a cell. The cell is the durable thing. The shape is a render.

7. Join SAIHM and try polymorphous cells in your own AI

If your current AI memory architecture has five boxes, polymorphous SAIHM cells collapse them into one. The fastest way to feel the difference is to try it on a real workload.

1. Join SAIHM at /join. Enrolment is a few clicks. PAYG and paid tiers available; see /pricing.
2. Connect your AI client to the SAIHM endpoint — one block of configuration, copy-paste from the quickstart page.
3. Paste the three prompts above. Save a fact one way. Recall it another way. Then a third way. Polymorphism is most convincing when you see it in your own workflow.

One protocol. One encryption posture. One audit trail. Every memory shape your AI Agent needs. Join SAIHM and see it for yourself.

Join SAIHM

Still have questions? The SAIHM chatbot is on every page (bottom-right). Ask it anything about polymorphous cells, recall, encryption, sharing, compliance, or pricing — you will get an immediate answer.