Lucio Durán
lucioduran.com

Projects / Agent Memory Lifecycle

Agent Memory Lifecycle

Auditable memory events with provenance, corrections, expiry, deletion and session-scoped lexical retrieval.

Author: · Python 3.11.13 · License: MIT

Page published: · Page updated:

Repository · Original evidence · Release v0.1.0

Problem and relevance to agents

Append-only memory may retrieve an obsolete fact after it has expired, been corrected or been deleted. Retrieval scores alone do not explain whether a fact should remain usable.

Explicit lifecycle decisions make memory selection inspectable before an agent consumes the selected facts.

Architecture

  1. Validate events. Assertions, corrections and tombstones carry source, session, recording time and validity bounds.
  2. Resolve before ranking. Filter session and validity, apply permanent corrections and deletions, and select one winning value per key before TF-IDF cosine retrieval.
  3. Explain and replay. Return conflicts, exclusions and provenance. Sanitized exports preserve replay structure while removing deleted payloads.

What this demonstrates

  • Deterministic correction and expiration semantics, including boundary timestamps.
  • Cross-session selection and deletion checks with explicit expected IDs.
  • Replayable traces showing why memory was selected or excluded.

Published results

Eighteen original fictional policy scenarios executed locally on a Xeon W-3223 using Python 3.11.13 and one thread.

Exact top-one ID set matches include empty expected results. Both methods filter by session. The baseline ignores corrections, expiry and deletion and uses simpler lexical ranking.

Lifecycle policy cases

Original evidence
MethodExact expected ID sets
Session-scoped append-only baseline8/18
Temporal policy and TF-IDF18/18

All policy scenarios passed, but the baseline comparison changes both ranking and lifecycle rules. It does not isolate a causal retrieval-quality gain. A single elapsed time is not a latency benchmark.

Reproduce

Python 3.11 or newer; replay and evaluation have no runtime dependencies.

Run measurements in a new destination and serialize heavy workloads. Within the original project series, use the shared benchmark_lock.py wrapper. Commands below use POSIX shell syntax.

Verify saved evidence

git clone https://github.com/lucioduran/agent-memory-lifecycle.git
cd agent-memory-lifecycle
git checkout d185070cadfd4d0b36e58d957ab7054ae90bdf18
python -m scripts.evaluate results/evaluation.json --verify

Repeat the local experiment

python -m scripts.evaluate results/new-run.json

Scope and limitations

  • The in-memory library assumes trusted callers and files. It provides no authentication, concurrency control, crash durability, encryption or secure erasure.
  • Deletion retains audit identifiers and timestamps and cannot erase external copies or old traces. Erased correction keys cannot be revalidated from sanitized exports.
  • Newer facts win by policy, not because they have been verified as true. Historical queries use current knowledge.
  • ASCII lexical retrieval and small author-known fixtures do not establish semantic understanding, production accuracy or security.

Sources and revision

The report identifies implementation commit 7b9a6b77399c38e461544c7897f812bab6e96ebd and a fixture hash, but contains no execution timestamp. The release and page dates are separate.