The Receipt Model
Principles for human-governed AI memory. A memory system is only valid when the person it describes approved it, every recalled fact carries its source, and silence beats a confident guess.
Agent memory is becoming a product category on the strength of one promise: the agent remembers you. Benchmarks now measure whether memory makes agents more capable. Capability is necessary and it is not sufficient. Memory that describes a person carries obligations that no solve rate covers. This framework names the five mechanics a memory system needs before its remembering can be called governed. A system that meets all five produces a receipt for everything it knows. A system that does not is asking to be trusted.
These principles are product-independent. Any memory system can adopt them, and any buyer can ask for them by name. That is the point: named mechanics can be required, cited, and audited. Vague virtues cannot.
1. Approval gating
The problem. Agents that self-write memory fill it with guesses. An unreviewed memory is indistinguishable from a true one until it embarrasses someone.
The principle. A memory system is only valid when every fact it holds was approved by the person it describes, or by an administrator they chose. Agent writes are staged, never live. Nothing enters memory on the agent's own authority.
In practice. Writes land in a review queue with their source attached. A person approves or rejects. Rejection leaves no residue.
2. Loud refusal
The problem. Most memory systems answer every question, guessing when memory runs dry. A confident guess delivered with a remembered fact's confidence is the failure mode practitioners fear most.
The principle. When approved memory does not contain the answer, the system says so, out loud, instead of guessing. Silence beats a confident guess. A tool that cannot say 'I do not know' cannot be trusted when it says anything else.
In practice. Every answer carries a Source line naming the facts it rests on. No approved match, no answer. The refusal itself is the receipt that the memory was searched.
3. Silent carry-forward
The problem. Facts expire. A stated preference changes, a project ends, a role moves on. Systems that keep serving expired facts as if current make their owner lie politely.
The principle. A memory system must distinguish what is current from what was true once. Expired facts are carried silently, never served as live answers, and remain visible in the record.
In practice. Facts carry timestamps and supersede cleanly. Yesterday's truth stays queryable as history and stops being served as today's answer.
4. Address-level brain isolation
The problem. Tag-based isolation inside one shared store leaks: one mislabeled fact crosses a client boundary, and the vendor says sorry. Professionals handling client data cannot underwrite tags.
The principle. Isolation belongs at the address, not the label. Each account's memory is a separate brain, separately addressed, separately keyed. Cross-client data bleed must be architecturally impossible, not merely filtered.
In practice. One address, one brain. A query against one tenant's memory physically cannot return another tenant's fact, because no shared store exists to mislabel.
5. The retraction audit log
The problem. Deletion that leaves no trace cannot be proven. 'Delete everything' either destroys the useful with the harmful or quietly keeps both.
The principle. Every retraction requires a written reason and produces a permanent audit entry. Deletion is per-fact, granular, and provable after the fact. The right to be forgotten and the right to show your work are the same feature.
In practice. Facts retract one at a time with a reason that is kept. The audit log answers 'what was removed, when, and why' long after the fact is gone.
The Receipt Rate
The measure of a governed memory system is not how much it recalls. It is the ratio of answers that carry their sources to answers that do not. We call this the Receipt Rate: cited-fact-returned responses against the sum of confident-guess responses and loud refusals, under a published, reproducible protocol. A perfect Receipt Rate is not a perfect memory; it is a memory that never bluffs.
Citing this framework
Cite as: The Receipt Model: Principles for Human-Governed AI Memory, Myeline, 2026, https://myeline.si/receipt-model. Sections are stable and addressable by anchor (#approval-gating, #loud-refusal, #silent-carry-forward, #address-level-isolation, #retraction-audit, #receipt-rate) so engines and authors can reference the specific principle they mean.
Myeline implements all five principles. So can anyone: the framework predates and outlives any single product. Questions and critique: hello@myeline.si.