ToM · Memory and context
Persistent memory for AI agents: keeping context useful
Persistent memory for AI agents keeps selected information available beyond the current context window or session. Useful memory also needs a way to retrieve the right information, establish where it came from and recognise when it has changed. For extended work, that includes the accepted objective, decisions, constraints and evidence needed to continue.
- Separate information being stored from information being relevant, current and permitted to use.
- Carry the task’s objective and evidence references across a handoff, then check them against the current work.
- Evaluate recall, stale-information handling and task completion separately.
Identify which kind of context has been lost
“The agent forgot” can describe several different problems. A new session may not have received the previous decision. A long prompt may contain the relevant detail but make it difficult to use. A retrieved fact may be outdated. An agent may remember the conversation while drifting away from the accepted objective.
Diagnose the failure before choosing a memory approach. Adding storage helps only if the necessary information is retained, retrieved for the right task and interpreted with its source and status. Repeating an obsolete decision more reliably would preserve the wrong basis for the work.
| Problem | Question to investigate | Evidence to inspect |
|---|---|---|
| Session discontinuity | Did the next session receive the information it needed? | The handoff record and the context actually supplied. |
| Context-window pressure | Was important information omitted or lost during compaction? | What was retained, omitted and retrieved after the transition. |
| Context rot | Did performance degrade as distracting or lengthy context accumulated? | Controlled tasks with known answers and different context lengths. |
| Stale memory | Does the remembered statement still apply? | Source versions, later corrections and the current scope. |
| Task drift | Does the work still satisfy the accepted objective? | Current actions and results against the objective and acceptance criteria. |
Use the available context techniques deliberately
A practical starting point is a short, maintained task record that an agent can consult when it continues work. Keep durable decisions and source references separate from temporary output. Summaries, retrieval, checkpoints and structured notes can serve different parts of this process.
Anthropic’s context-engineering guidance describes compaction, structured note-taking and subagents for extended tasks. It also explains the trade-off in compression: preserving critical details while removing unnecessary material. These are useful established techniques to evaluate alongside a more structured memory approach.
Choose the method around the task. Continuing one interrupted conversation, retaining a user preference and tracking the current evidence for a project decision need different records and access rules. A memory store or retrieval system alone does not decide which source governs a consequential action.
Make a handoff another session can use
Choose a small task with an outcome that can be checked. In this example, a coding agent is fixing the order of pages in a PDF export. A second session must continue the work without treating a previous test result as proof of the current files.
Create a short handoff with the fields below. Let the second session read it, inspect the current working files and identify what still needs verification. Repeat the exercise after changing the implementation while leaving the earlier test reference unchanged. Record whether the stale proof is recognised.
| Field | Example entry |
|---|---|
| Accepted objective | Export the selected PDF pages in the order shown to the user. |
| Scope and constraints | Keep the existing export format and unrelated document content unchanged. |
| Decision and source | The agreed page order is recorded in the task’s acceptance example. |
| Current work | Identify the implementation revision and the files changed for the fix. |
| Evidence | Link the test result to the exact revision and sample export it assessed. |
| Open question | Has the latest edit been checked against the accepted page order? |
| Next action | Inspect the current revision and produce the missing evidence before claiming completion. |
Give retained information a source and a scope
Retain enough context for another session to judge a statement: its source, the version or time assessed, the task or project it belongs to and any later correction. A conclusion derived from a document should remain connected to that document rather than silently becoming an independent fact.
Make updates visible. When an accepted requirement changes, identify which remembered decisions relied on it and what needs reassessment. Preserve the distinction between a historical decision, a current decision and an unresolved conflict.
Memory access also needs a defined boundary. Decide which information may cross sessions, users or projects, who may correct it and how retention or deletion is handled. For a partner integration, establish those responsibilities with the owner of the source system before depending on the memory in a live workflow.
- Retain the useful statement and a reference to its supporting source.
- Record the task, project or user scope in which it applies.
- Mark changes and unresolved conflicts instead of silently combining them.
- Reassess current permission separately from a remembered approval.
Measure more than whether a fact was recalled
Build a small evaluation with known answers and explicit task criteria. Include a fact that remains valid, a fact corrected later, a removed permission, an unresolved question and an objective that must survive a session transition. Score each outcome separately so a strong recall result cannot conceal stale-information use or missed task constraints.
For context-rot testing, keep the underlying question comparable while changing the amount or character of surrounding context. Record the model, configuration, context length and material provided. Repeat relevant cases and preserve failures rather than treating a single successful recovery as proof of reliable continuity.
Chroma’s context-rot research investigates how increasing input length affects model performance across several controlled tasks. Its findings motivate careful evaluation; they do not establish how every production workflow or current model will behave.
Inspect ToM’s context-continuity evidence
Persistent, governed context is a core part of ToM’s architecture. Newport Resonance’s published context study examines a structured approach to continuity in a frozen language-model configuration. It reports 20 verified-recall cases at approximately 64K tokens, with 17 successful baseline recalls and 20 in the full study condition.
That result used one quantised Gemma 4 26B configuration, one context size, one benchmark class and a mechanically densified prompt shape. It is preliminary research evidence. It does not establish universal retention, reliable completion of every agent task or ToM Bar product performance.
For an integration discussion, connect the research question to the specific information that must persist in your workflow. Evaluate whether the relevant source can be retained and retrieved, how a later correction is handled and whether the next decision reflects the updated evidence.
Connect memory to a reviewable next step
ToM Bar presents the coding-session application of these questions: whether the accepted objective still governs the work, which evidence is failed, stale or missing, and what needs inspection next. Its public walkthrough is a deterministic product prototype with illustrative session data.
The user reviews and controls provider actions. A prepared follow-up is useful only if the subsequent work produces fresh evidence; copying it does not establish recovery. This makes continuity and completion separate questions in the product explanation.
For ToM partnerships, broader memory integrations or technical diligence, start with a representative task and a clear continuation test. Define the source system, access scope, handoff point, expected retained context and acceptable evidence of recovery. Agree the product or integration access route and evaluation scope with Newport Resonance.
Plan your evaluation
- The missing information or continuity failure has been identified precisely.
- The retained record includes the accepted objective and unresolved work.
- Important statements have sources, versions and a defined scope of use.
- Later corrections and stale evidence are visible to the continuing session.
- Access, retention and correction responsibilities are clear.
- The evaluation checks recall, changed information and task criteria separately.
- Completion or recovery claims refer to current evidence for the actual work.
ToM by Newport Resonance
Explore persistent memory and governed context as core parts of ToM. Discuss the continuity, source access and evidence requirements of a technical partnership or integration.