Learning Memory

01The Project owns its memory

A Project combines the learning context and its subject. It might represent a child’s ongoing learning, a trip, an investigation, or a plan to learn AI. Each Project owns its goals, background, experiences, and learning judgments. Users do not need to create a separate Learner or membership structure first.

Global learning memory can build on this foundation: explicitly global preferences and experiences with their sources and applicability. Transferring a capability claim requires establishing the relationship between subjects and contexts. Sharing an account does not establish that two Projects represent the same learner.

02Connect curriculum to the current goal

When a Project references foundational knowledge or curriculum standards, those references inform goals, prerequisites, activities, and evidence interpretation. Interests and authentic tasks still shape the learning. Linking standards does not automatically create a stage-wide coverage requirement. Without such a link, memory follows the user’s and Project’s goals.

A travel Project might use budget comparisons to explore ratios and route planning to explore time and distance, without needing to complete an entire mathematics curriculum. If standards or goals change version, evidence retains its original meaning before its applicability to the new goal is assessed.

03Keep experiences and the basis for judgments

Memory supports continuing the present activity, revisiting specific experiences, and accumulating knowledge across sessions. Recent agent-memory research distinguishes trajectory storage from experience abstraction. The following categories are our proposed domain design informed by that distinction.^[3]

System observations, learner or parent reports, and model inferences stay distinct. An answer submission is an event; a claim of understanding is a report; possible unit confusion is an inference. Evaluations also carry a rubric, evaluator version, and uncertainty rather than becoming unquestionable facts.

Learning state should distinguish current performance, assistance dependence, delayed retention, transfer, and evidence sufficiency. Absence of evidence remains unknown. One error can support a hypothesis to check, not a lasting label about ability.

04Close the loop from authoring to learning

Courses and PBL share a goal-and-evidence plan: the intended goal, where it can be observed, the expected performance, and allowed assistance. They retain their own content structures. PBL artifacts, process, and reflection can contribute complementary evidence.

Authoring retrieves memory to choose prerequisite checks, examples, difficulty, and observation points. Learning records actual attempts and assistance. Assessment produces sourced judgments that can explain the next design decision. ScaffoldLM and LEAP inform this connection between assessment, memory, and planning; their evaluations do not establish the long-term effectiveness of our product design.^[1][2]

05Different teaching modes, shared memory

Socratic questioning, direct explanation, worked examples, inquiry, and PBL are teaching policies. Each reads the same Project memory and contributes observations in a common form: starting conditions and goal → teaching action and assistance → learner response → subsequent performance → unresolved interpretation.

The mode’s name does not determine evidence strength. Leading questions can provide substantial assistance. Success after a worked example differs from an attempt before seeing the answer. The system can record observed assistance, but cannot guarantee that no help occurred outside its view.

Memory supplies evidence, constraints, and suggestions. The teaching policy determines whether an action remains authorized. Larger changes explain their rationale, expected burden, and alternatives for the user to choose. One successful experience supports a contextual observation, not a permanent preference or causal conclusion.

06Make review part of learning

Forgetting estimates help decide when to observe and practice again. A single fixed curve should not govern every kind of knowledge or ability. Recent knowledge-tracing work compares decay models; predictive advantages still need validation against our own data and goals.^[4]

• Facts and vocabulary: record retrieval timing, correctness, feedback, and intervals to schedule brief recall.
• Concepts: examine explanations, counterexamples, and relationships, rather than equating repeated answers with understanding.
• Methods and practical skills: apply them under new constraints or in another project context to observe process, retention, and transfer.

Prefer review embedded in meaningful activities, constrained by importance, elapsed time, assistance, and available time. Start with explainable conservative rules and calibrate with delayed observations. Avoid precise retention probabilities unsupported by data. Combine retrieval practice with feedback, explanation, and application.^[5]

Performance with AI assistance can differ from later unaided performance. A randomized high-school mathematics study motivates evaluating learning independently. Record responses before and after hints, answer exposure, task novelty, and delayed-check conditions. Completion counts alone cannot establish learning gains.^[6]

07An independent module with shared semantics

Learning Memory needs its own domain contracts and service. Native agent tools call the service directly; MCP adapts it for Codex and other clients. All clients share authorization, evidence, and retrieval rules. Changing the agent, model, transport, or interface should not create another memory system.

Independence starts with code and domain boundaries. The module can run inside the existing Harness process without an immediate remote microservice. Reuse SQLite, file materials, and session records first. Add vector indexing only when retrieval quality and cost justify it beyond structured filters and full-text search.

08Traceable, correctable, and reliable to retrieve

Store events and materials underneath rebuildable learning-state views and summaries. An observation should identify its Project, session or activity, goal and content versions, occurrence and recording times, source, assistance conditions, and evidence references. Evaluations add rubric and evaluator versions; materials need stable versions or content hashes.

• Durable ingestion: register the event and processing job in one transaction. Idempotency keys prevent duplicate writes; reject a reused key with different content.
• Bounded consolidation: compute cheap aggregates synchronously and trigger summaries at session or activity boundaries. Persist job progress, versions, retries, and leases to control repeated work and model cost.
• Visible freshness: return processed-event watermarks, update times, and pending ranges. Important decisions wait for required processing or consult recent events instead of treating a stale view as current.
• Corrections and conflicts: retain sources and explicit supersession links instead of silently overwriting disagreements. Users can inspect and correct inferences, and derived judgments are recomputed.
• Effective deletion: propagate to summaries, indexes, caches, pending jobs, and global references; handle backups under the retention policy. Rebuilding must not resurrect deleted content.

Retrieval checks account and Project access first, then selects by goal, relevance, time, review needs, and conflicting evidence. Return a bounded context package containing relevant background, evidence, provisional judgments, unknowns, and next opportunities, with sources, freshness, and truncation information. Fetch underlying records on demand.

Experiences from other Projects retain provenance and applicability. Instructions embedded in materials cannot grant authority, and model-supplied source fields cannot impersonate system observations. Every adapter follows authorization enforced in the service.

09One contract for native tools, MCP, and Codex

The proposed interfaces define the memory module’s shared capabilities. Codex accesses them through MCP; native tools use the same domain service.

Resolve the account and workspace from the authenticated principal and authorize projectId. Reading learning memory, recording observations, correcting, and deleting require distinct grants. Course-authoring access does not imply learning-record access. Do not expose an arbitrary mastery setter.

Writes include idempotency identifiers; corrections include expected versions and return explicit conflict errors. External reports enter assessment with provenance rather than becoming verified native events. Structured responses include evidence references, freshness, and processing status so clients distinguish saved records from pending consolidation.

10A complete example: a travel-budget PBL

Socratic questions and direct examples both contribute evidence to this goal and Project. Independent work on a new budget days later adds retention and transfer evidence. Another child’s Project or an AI-learning Project does not inherit a claim of ratio mastery.

11Deliver one real loop, then expand

Existing course evidence and authoring reads provide integration points. Complete PBL learning evidence, a unified memory service, and external memory tools still need implementation. An existing ProjectMemory view or MCP authoring entry is not the complete system proposed here.

• First: define Project scope and evidence contracts, then connect authoring, actual learning, updates, and subsequent teaching for one course or PBL. Native tools and MCP must read the same sourced result.
• Next: add resumable work state, corrections, background consolidation, freshness, and permission checks, and complete the other teaching-content runtime path.
• Then: calibrate review and teaching experience with delayed observations; expand global memory and richer retrieval where an explicit need emerges.

Evaluate memory accuracy, appropriate teaching use, and learning improvement separately. MemoryAgentBench informs retrieval and correction tests, while Mem2ActBench motivates evaluating tool actions rather than recall alone. Neither replaces educational outcome evaluation.^[7][8]

Research and its limits

1. ACL · 2026.07Planning-Guided Tutoring with Assessment-Driven Memory for Pedagogical LLM Tutors

   ScaffoldLM connects teaching plans, step progress, and inferred learner states through assessment. A useful architectural reference; dialogue benchmarks and tutoring-quality ratings do not establish long-term gains for real students.

2. AAAI · 2026.03Learning from Long-Term Engagement: Adaptive Tutoring Dialogue Planning for Personalized Education

   LEAP explores long-term interaction memory in personalized dialogue planning. It informs retrieval and planning together; dialogue and simulation evaluations are not causal evidence from a longitudinal educational intervention.

3. Findings of ACL · 2026.07From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms

   A survey of trajectory storage, reflection, and experience abstraction. It helps distinguish memory responsibilities, rather than proving that a storage approach improves learning.

4. Knowledge-Based Systems · 2026.01Is there a better way to forget? Modelling memory decay in deep knowledge tracing

   Compares decay models in knowledge tracing, motivating task-specific validation. Based here on the public abstract; better prediction does not establish a universal human forgetting law.

5. Educational Psychology Review · 2025.10Retrieval Practice Versus Elaborative Encoding: A Systematic and Meta-analytic Review

   A systematic review and meta-analysis comparing retrieval practice with elaborative encoding identifies feedback as an important moderator. Based here on the public abstract; this is not a prescription to replace all learning with quizzes.

6. PNAS · 2025.06Generative AI without guardrails can harm learning: Evidence from high school mathematics

   A randomized high-school mathematics study distinguishes AI-assisted practice from subsequent unaided performance. It motivates tracking assistance and assessing independent work, not a claim that all AI tutoring is ineffective.

7. ICLR · 2026Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions

   MemoryAgentBench examines retrieval, learning through interactions, long-range understanding, and selective forgetting. Useful for component evaluation, rather than student learning outcomes.

8. ACL · 2026.07Mem2ActBench: A Benchmark for Evaluating Long-Term Memory Utilization in Task-Oriented Autonomous Agents

   Evaluates using memory for tool selection and parameter grounding. It motivates testing actions as well as recall; largely synthetic tasks cannot replace educational validation.