What does this skill do, and when should you use it?
This skill is for agent systems that must retain knowledge across sessions, track entities, and reason over accumulated information. It guides teams from filesystem-backed memory toward vector stores, graph memory, and temporal knowledge graphs based on retrieval needs. It covers framework selection, retrieval strategies, consolidation, stale or conflicting facts, and benchmark choice. It is an architectural guide with examples, not a standalone memory runtime.
Compares Mem0, Zep/Graphiti, Letta, Cognee, LangMem, and filesystem memory by retrieval pattern and trade-off; maps working, short-term, long-term, entity, and temporal-KG layers to implementation choices; explains semantic, entity-based, temporal, and hybrid retrieval; recommends consolidation and recovery handling for empty, stale, conflicting, or failed memory operations; identifies LoCoMo, LongMemEval, and DMR as evaluation options; and provides Python examples for Mem0, temporal queries, and Cognee ingestion and search.
- An agent engineer is building cross-session memory for user preferences, domain knowledge, or entity identity.
- A production team needs to decide when to move from files to vector or graph-backed memory.
- A system must represent changing facts, validity intervals, or historical states.
- An architecture team is comparing Mem0, Zep/Graphiti, Letta, Cognee, and LangMem.
- An evaluation team needs to measure long-conversation retention and retrieval quality with published benchmarks.
How do you install this skill?
- Do not treat scripts/memory_store.py as production persistence: it uses pseudo-embeddings and in-memory structures, leaves consolidate() empty, and provides no deletion, access control, encryption, or rollback mechanism.
- Treat benchmark figures and framework capabilities as leads requiring revalidation; the supplied material lacks independent results for this skill's key paths.
- Before persisting user memories, add consent, data classification, tenant isolation, retention, deletion-rights, and conflict-audit policies.
- Pin Python and third-party dependency versions before deployment, and verify the frameworks' actual APIs, cloud restrictions, and mainland-China reachability.
- Local filesystem
Copy the complete skills/memory-systems directory into the host's Agent Skills directory. For a Codex project, for example: mkdir -p .codex/skills && cp -R skills/memory-systems .codex/skills/. Preserve the directory layout rather than flattening SKILL.md; the repository also documents installation as a Claude Code plugin marketplace containing the collection.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/muratcankoylan/Agent-Skills-for-Context-Engineering.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/memory-systems" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Design cross-session memory for an agent that tracks changing user preferences, and compare filesystem memory, Mem0, and Zep/Graphiti.
Use it in an Agent Skills-compatible host with a request such as: “Design cross-session memory for an agent that tracks changing user preferences, and compare filesystem memory, Mem0, and Zep/Graphiti.” It is intended to activate for persistent semantic memory, entity tracking, temporal validity, retrieval design, consolidation, and memory benchmarking. Complete runtime integration and dependency-installation instructions are not provided by this skill.
What are this skill's strengths and limitations?
- Offers a clear escalation path from filesystem memory to temporal knowledge graphs.
- Covers selection, retrieval, consolidation, conflict handling, and failure recovery.
- Treats benchmark results as dated, source-specific signals rather than absolute rankings.
- Includes concrete Mem0, temporal-query, and Cognee examples.
- Does not provide a standalone executable implementation or complete installation guide.
- Framework and benchmark claims should be rechecked before product decisions.
- No skill-specific test suite or platform validation is shown.
- Graph and advanced memory frameworks add infrastructure and operational complexity.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Persistent Semantic Memory Design this page | 53 · Use with care | ★ 18k | 16d ago | MIT |
| SkillOpt-Sleep ✓ Microsoft · Official | 64 · Recommended | ★ 18k | 22d ago | MIT |
| SkillOpt-Sleep: Sleep-Cycle Self-Improvement for Your Local Claude Agent ✓ Microsoft · Official | 64 · Recommended | ★ 18k | 22d ago | MIT |
| NeMo Evaluator Plugin Skill ✓ NVIDIA · Official | 58 · Recommended | ★ 3.6k | 14d ago | Apache-2.0 |
| Skill Feedback Writer | 37 · Not recommended | ★ 32k | 15d ago | Apache-2.0 |
The skill explicitly compares Mem0, Zep/Graphiti, Letta, Cognee, LangMem, and filesystem memory. Its central recommendation is to start with the shallowest viable layer and add structure only when retrieval quality or reasoning requirements justify it: filesystem memory for simple agents and prototypes, Mem0 for fast production adoption, Zep/Graphiti for temporal and relationship modeling, Letta for deep introspection, and Cognee for dense customizable semantic graphs.
How did FollowSkills review this skill?
The skill explicitly recommends privacy consideration, retention policies, deletion rights, session scoping, and non-blocking storage-failure handling. However, it does not define least privilege, user confirmation, sensitive-data classes, encryption, access control, data-flow disclosure, or rollback for persisted data, so 9 points are deducted. No malicious behavior or obvious overreach is shown.
The document is internally organized and covers empty retrieval, stale results, conflicting facts, and storage failures, with an accompanying implementation script. However, consolidation is unimplemented, dependencies are not pinned, hash-based pseudo-embeddings are not stable across processes, and the script's text-field behavior differs from some examples and reference code. Because this is static review with no skill-specific tests, the score is conservatively capped at 6, deducting for reproducibility and incomplete failure diagnostics.
Activation scenarios, intended use, and boundaries with filesystem-context, context-compression, and context-optimization are clearly stated, covering file, vector, graph, and temporal memory. It lacks sufficiently explicit non-fit ranges, input/output contracts, platform requirements, Chinese-language support, and mainland-China network reachability, so 4 points are deducted.
The skill uses progressive structure with frontmatter, activation criteria, layered concepts, practical guidance, examples, error recovery, integration notes, and references. The repository supplies an MIT license, version, creation/update dates, and author information. It lacks a clear maintenance owner, update procedure, changelog, dependency-installation guidance, FAQ, and a fully demonstrated path for validating internal references, so 5 points are deducted.
The skill provides directly usable architecture-selection, retrieval, temporal-modeling, and recovery guidance, plus Python examples and an integration script, so it plausibly completes the core design task. However, the examples use pseudo-embeddings, consolidation is a no-op in the integrated implementation, several components are explicitly demonstrative, and representative outputs or comparative benefits are not verified. Under the static cap, it receives 6, with deductions for limited production readiness and evidence.
The repository includes reference lists, claimed benchmarks, router regression materials, and CI, but the supplied CI primarily validates researcher tooling and other examples rather than the memory-systems key paths. Benchmark figures and external papers are not supplied with reproducible result artifacts in this material. This supports limited auditability but not independent reproduction, so 4 points are awarded with deductions for thin cross-validation and skill-level test coverage.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →