Agent Memory Errors Handbook: Cases 11–15

This installment is part of the Agent Memory Errors: 100 Cases Handbook series.

Case 11: Long-term Memory Management Errors — Episodic Memory Duplicate Storage

Scenario

After every conversation, the Agent stores a dialogue summary in long-term memory. A user repeatedly asks the same question, such as “What is my server password?” Each answer creates a new memory: “User asked for server password; answer is XXX.” After six months, 47 nearly identical memories accumulate under this topic.

Error Manifestation

When answering related questions, the Agent retrieves several of these memories. Because the summaries differ slightly in wording, the Agent gives inconsistent answers across sessions, sometimes citing multiple versions in one reply and appearing self-contradictory.

Solution

1. Deduplicate semantically before storing: compare new memories with existing ones and merge when similarity exceeds a threshold
2. Track “last updated” timestamps and access counts; merge by keeping the latest version and accumulating counts
3. Periodically run memory compaction: cluster memories on the same topic and generate a unified summary

Principle

Continuous accumulation of episodic memory causes storage bloat and redundancy. Human memory consolidates similar experiences; Agent memory systems need analogous mechanisms to stay consistent.

Case 12: Long-term Memory Management Errors — Procedural Memory Overwrite Bugfix

Scenario

A programming Agent stores procedural memory for “How to fix Python’s UnicodeDecodeError.” A user encounters a CSV encoding error, and the Agent generates a scenario-specific fix and stores it. Later, another user encounters a different UnicodeDecodeError (JSON encoding), the Agent generates a new fix, and overwrites the old memory.

Error Manifestation

When the first user hits the CSV problem again, the Agent gives the JSON-scenario fix (using json.loads’s encoding parameter), which is completely inapplicable and confuses the user.

Solution

1. Store procedural memory as condition-action pairs, explicitly recording each memory’s applicable scenario
2. Match conditions before overwriting: only overwrite when the new solution applies to the same or a more general scenario
3. Add memory version control so historical versions remain available and the most relevant one can be selected

Principle

Procedural memory encodes how-to knowledge and is tightly coupled to specific scenarios. Direct overwriting destroys the scenario-solution mapping. Conditional storage lets knowledge for different scenarios coexist.

Case 13: Long-term Memory Management Errors — Semantic Memory Inheritance Conflict

Scenario

A knowledge-graph Agent builds semantic memory with an object-oriented multiple-inheritance model. The entity “Electric Vehicle” inherits from “Car” (attribute: needs refueling) and “Electric Device” (attribute: needs charging). When the user asks, “Does an electric vehicle need refueling?” the Agent must choose between two contradictory parent attributes.

Error Manifestation

The Agent gives inconsistent answers: once “needs refueling” (from Car), once “no” (from Electric Device), with no consistent decision rule.

Solution

1. Set inheritance priorities (e.g., subclass > direct parent > indirect parent) or use C3 linearization
2. Explicitly flag conflicting attributes and trigger a conflict-resolution flow that asks for more context
3. Use probabilistic inheritance: compute attribute likelihood weighted by association strength to each category

Principle

Inheritance conflicts in semantic memory are a classic knowledge-representation problem. Without conflict resolution, systems behave nondeterministically and erode user trust.

Case 14: Long-term Memory Management Errors — Memory Access Pattern Hotspot Exhaustion

Scenario

A social-media Agent maintains a user-profile memory for every followed user. A celebrity has one million followers, and every follower’s query about “how to interact with X” triggers retrieval of that celebrity’s profile. The memory becomes a hotspot, accessed thousands of times per second and exceeding the vector database’s read quota.

Error Manifestation

After rate limiting kicks in, later queries for that profile time out or degrade to empty results. The Agent cannot offer personalized suggestions, and users complain about instability.

Solution

1. Add multi-level caching for hot memories: L1 in-process cache → L2 Redis → L3 vector database
2. Shard read replicas of hot memories to distribute load
3. Use approximate-memory snapshots: periodically generate static summary files (e.g., JSON) for hot profiles and serve them from object storage

Principle

Every storage system has access limits. Frequently accessed long-term memory concentrates load on a single point. Multi-level caching and read sharding are standard patterns for hotspots.

Case 15: Long-term Memory Management Errors — Forgetting Curve Misalignment

Scenario

The Agent applies the Ebbinghaus forgetting curve: new memories decay 20% after one day, 50% after seven days, and 80% after thirty days. In a project-management scenario, the user creates a memory: “Product launch plan in three months.” Under the default curve, the memory decays to near-unretrievable before the launch.

Error Manifestation

A week before launch, the user asks, “What is our launch plan?” The Agent cannot retrieve it from long-term memory and replies, “I have no information about the launch plan.” The user thinks the Agent has a bad memory.

Solution

1. Use custom decay strategies by memory type: different curves for todos, plans, and facts
2. Introduce importance markers: users or the system can mark a memory as important, skipping decay or extending retention
3. Set reminders for time-sensitive memories so their weights rise automatically at trigger times

Principle

A single forgetting curve cannot fit every business timescale. Project management needs month-level memory; casual chat may need only hour-level memory. Adaptive decay matches business needs.

Source: Agent Search