How Do You Know Which Memories Are Still Accurate

Confidence Score as an Accuracy Indicator

The confidence score is the single best indicator of memory accuracy. In Adaptive Recall, confidence ranges from 0 to 10, starting at 5.0 for new memories. The score increases when independent sources corroborate the memory and decreases when contradictions are detected. A memory with confidence 9.0 has been confirmed multiple times across different contexts, which makes it highly likely to be accurate. A memory with confidence 3.0 has either been contradicted or was never corroborated, which means its accuracy is uncertain.

Confidence is a better accuracy indicator than age because it reflects evidence rather than assumptions. A two-year-old memory with confidence 9.5 is almost certainly accurate because it has been confirmed repeatedly over its lifetime. A two-day-old memory with confidence 5.0 might be accurate but has no corroborating evidence yet. Confidence tracks the system's actual evidence about each memory's reliability.

Corroboration Count

Corroboration count measures how many independent observations have confirmed the same fact. A memory that says "the database runs on PostgreSQL 15" and has been corroborated by three separate conversations is more likely accurate than one that was stored once and never confirmed. Each corroboration is an independent data point that the fact is correct.

Corroboration is particularly valuable for detecting memories that were wrong from the start. If a memory was stored based on a misunderstanding and the correct information is stored later, the consolidation process detects the contradiction. But if the original memory was simply never corroborated while the correct version is confirmed multiple times, the corroboration count alone provides strong evidence about which is more reliable.

Contradiction History

When the consolidation process detects that a memory contradicts newer information, it records this in the memory's metadata. Memories that have been involved in contradiction resolutions carry a history that indicates they may contain information that was once true but has been superseded. Even if the contradiction was resolved in the memory's favor, the fact that it was challenged provides useful context about its reliability.

Memories with no contradiction history are not guaranteed accurate, but they have at least never been directly challenged by newer information. Combined with high confidence and recent corroboration, a clean contradiction history is a strong indicator of accuracy.

Access Patterns as a Signal

Memories that are regularly retrieved and used by applications or users are more likely to still be accurate, because inaccurate memories tend to be noticed and corrected when they surface in actual use. A coding assistant memory about "use environment variables for configuration" that gets retrieved every week and never triggers a correction is probably still accurate. A memory about a specific API endpoint that has not been retrieved in three months may have been superseded by an endpoint change that the system has not yet learned about.

This signal is imperfect because some accurate memories are rarely retrieved (foundational facts that are not directly queried), and some inaccurate memories may be retrieved without the error being caught. But in aggregate, access patterns correlate with accuracy because active use creates opportunities for error detection and correction.

Combining Signals

No single signal is sufficient to determine accuracy. The strongest assessment combines all four. A memory with high confidence, multiple corroborations, no contradiction history, and recent active retrieval is very likely accurate. A memory with low confidence, no corroborations, a past contradiction, and no recent access is likely stale or inaccurate.

Adaptive Recall uses these combined signals in its cognitive scoring model. During retrieval, memories with higher confidence and stronger corroboration rank above memories with weaker signals, even if the text similarity scores are comparable. This means that retrieval naturally surfaces the most reliable information first, without requiring the caller to manually filter or verify results.

When Automated Checks Are Not Enough

Automated accuracy assessment has limits. It cannot verify facts against external reality, only against other memories in the same store. If an entire domain's understanding shifts and no contradicting memory has been stored, the system has no way to detect that its existing memories are outdated. For critical applications, periodic human review of low-confidence and old, uncorroborated memories is a valuable supplement to automated lifecycle management.