Appendix B — References and Reproducible Resources

This appendix lists only sources that can be independently accessed and recomputed. Every number cited in the body text can find its origin here; sources that cannot be listed here are not cited in the body text.

All links and version information are as of 30 October 2026. Product documentation and repositories are updated continuously; if you find a version difference when recomputing, the official page at that time prevails.

B.1 Academic papers

ID	Title	Authors / venue	Relation to this book
arXiv:2605.06716	From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms	Jinghao Luo et al., ACL 2026 Findings, submitted 2026-05-07	Source of the three-stage evolution framework (storage / reflection / experience); the backbone of Chapter 19
arXiv:2410.10813	LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory	Di Wu et al., ICLR 2025	A composite benchmark of 500 questions over five abilities; the original source of "roughly a 30% accuracy drop in sustained interaction"
arXiv:2504.19413	Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory	Prateek Chhikara et al.	LOCOMO comparison: 26% relative improvement in LLM-as-a-Judge, 91% lower p95 latency, over 90% token cost saving
arXiv:2501.13956	Temporal knowledge graph methods in Zep / Graphiti	Zep team	Methodological source for the dual time axis and incremental graph updates

On the accurate formulation of the three-stage framework: the survey formalizes the evolution as storage (trajectory preservation), reflection (trajectory refinement), and experience (trajectory abstraction), and analyses three drivers — the necessity of long-range consistency, the challenge of dynamic environments, and the goal of continual learning. In the frontier "experience" stage it focuses on two mechanisms: proactive exploration and cross-trajectory abstraction. This book does not extrapolate beyond that framework.

B.2 Products and implementations

B.2.1 Mnemosyne OS

Resource	URL	Notes
PyPI	https://pypi.org/project/mnemosyne-os/	Package name mnemosyne-os, current version 8.1.0.post1, MIT
Source repository	https://github.com/MnemosyneOS/mnemosyne	Main branch contains mnemosyne/, storage/, security/, scripts/, docs/
Benchmark repository	https://github.com/MnemosyneOS/mnemosyne-bench	Unified scorer, three-method adapters, naive baseline, CN-Scenarios dataset
Hosted console	https://ai-memory.net	Registration, API key management, memory browsing, subscription management
Compatibility notes	docs/COMPATIBILITY.md in the repository	Mapping of the conventional agent-memory call shape
Deployment notes	docs/DEPLOY_DEEPSEEK_HARNESS.md in the repository	MCP deployment walkthrough

Installation and minimal reproduction:

pip install mnemosyne-os
python -c "from mnemosyne import Memory; m=Memory(); m.add('I prefer dark mode', user_id='alice'); print(m.search('preference', filters={'user_id':'alice'})['results'][0]['memory'])"

Self-check scripts (in the repository; clone first):

python verify.py
python scripts/verify_api.py
python scripts/verify_precision_recall.py
python scripts/verify_recall_quality.py
python scripts/verify_memory_lifecycle.py --brain-dir ./mem --src-root .
python scripts/verify_formation.py
python scripts/verify_graph_health.py
python scripts/verify_l1_l3.py
python scripts/verify_v810.py

Benchmark reproduction:

git clone https://github.com/MnemosyneOS/mnemosyne-bench
cd mnemosyne-bench
python -m harness.run --systems mnemosyne,naive --datasets longmemeval
python scripts/verify_harness.py

A fact that must be stated alongside the above: the benchmark repository's results/2026-10-24.md labels itself a seed result table, whose purpose is to carry the scores already published in the main repository's README so the two never drift; the only machine-readable run record currently committed is a naive baseline run on CN-Scenarios (hit@5 = 0.0, 6 questions). The README figures — LongMemEval 96.2, LoCoMo 94.8, BEAM 1M 68.5, BEAM 10M 53.9 — should therefore be cited as claims pending recomputation by a complete in-repository run, not as verified results.

B.2.2 Mem0

Resource	URL	Notes
Research page	https://mem0.ai/research	The four algorithmic changes (single-pass ADD-only extraction, multi-signal retrieval, temporal reasoning, memory decay) and token comparison data
Documentation	https://docs.mem0.ai/	Separate sections for open source and platform; default components, self-hosted stack, component configuration
Pricing	https://mem0.ai/pricing	Free and Custom tiers
Repository	https://github.com/mem0ai/mem0	Open-source library and self-hosted server
Benchmark data	https://github.com/mem0ai/memory-benchmarks	The authoritative source for published scores

Scope note: Mem0's official research page in October 2026 states "single-pass ADD-only extraction — extraction collapses to one LLM call that only adds, with no overwrite or delete step, so history survives intact with about 2x faster extraction"; and "memory decay — recently accessed memories get a soft ranking boost while idle ones dampen over time at zero search latency cost. Nothing is ever deleted or hidden." This book describes accordingly, and does not cite any unpublished decay coefficient.

Benchmark scores must carry three scope markers (Platform / OSS, Top 50 / Top 200, accuracy / pass rate). The official figures are in the mem0ai/memory-benchmarks repository README, not on the mem0.ai/research page (that page's four charts are client-rendered and capture as 0.0, so they must not be cited):

Benchmark	Top 200	Top 50	Metric
LongMemEval (500 questions)	94.4% (472/500)	94.8% (474/500)	Accuracy
LoCoMo (1540 questions)	92.5% (1425/1540)	91.8% (1414/1540)	Accuracy
BEAM 1M (700 questions)	70.1% (491/700)	67.1% (470/700)	Pass rate
BEAM 10M (200 questions)	50.5% (101/200)	45.5% (91/200)	Pass rate

BEAM's pass rate and LoCoMo's accuracy are not directly comparable. The OSS self-hosted configuration has a separate set (LongMemEval overall 91.0%, GPT-5 answering and judging), which is not the same set as the Platform figures.

B.2.3 Zep / Graphiti

Resource	URL	Notes
Zep documentation	https://help.getzep.com/	v4 is the default version; a /llms.txt index is available
Context MCP Server	https://api.getzep.com/mcp	Exposes 12 tools: search_graph, get_user_summary, get_subgraph, get_node_neighbors, list_episodes, add_memory, list_graphs, search_graph_in, get_subgraph_in, get_node_neighbors_in, list_episodes_in, add_memory_to_graph, plus a zep://graphs/directory resource
Graphiti repository	https://github.com/getzep/graphiti	Apache-2.0 framework for temporal context graphs, with hybrid retrieval
Pricing and plans	https://www.getzep.com/pricing	MCP billed per seat; custom OIDC requires Enterprise

Scope note: Zep positions itself as "the unified context layer for enterprise data", organizing data in temporal Context Graphs and serving Context Lake through the Konig graph database service. The Context MCP Server authenticates the end user rather than a Zep account member, signing in through the customer's own identity provider.

B.2.4 Letta

Resource	URL	Notes
Documentation	https://docs.letta.com/	Concepts include Stateful agents, Conversations, MemFS, shared memory
Current source	https://github.com/letta-ai/letta-code	Contains the agent harness, terminal UI, App Server, channels, runtime
Historical source	https://github.com/letta-ai/letta (archive branch)	The retired V1 API server

Scope note: Letta, formerly MemGPT, distributes as the npm package @letta-ai/letta-code, with Letta Cloud and a TypeScript Agent SDK. This book does not assert the current implementation status of "sleep-time compute"; it is cited only as an architectural idea.

B.2.5 LangMem

Resource	URL	Notes
Documentation	https://langchain-ai.github.io/langmem/	Conceptual guide: semantic / episodic / procedural memory; conscious and subconscious formation; memory namespace
Repository	https://github.com/langchain-ai/langmem	Memory SDK; persistence delegated to a LangGraph store

B.2.6 Cloudflare Agents

Resource	URL	Notes
Documentation	https://developers.cloudflare.com/agents/	Durable agent runtime; conversation state and memory, tools, Agent Skills, workflows, file system, sandbox, MCP
Documentation index	https://developers.cloudflare.com/agents/llms.txt	Full-site page index

B.3 Reproducible observation methods

The data in this book's "Technical Observations" chapter can be recomputed with the following queries. Both match only arXiv metadata fields and therefore return lower bounds.

https://export.arxiv.org/api/query?search_query=all:%22agent+memory%22&max_results=1
https://export.arxiv.org/api/query?search_query=abs:%22memory%22+AND+abs:%22LLM+agent%22&max_results=1

Measured on 30 October 2026, the opensearch:totalResults values were 600 and 841 respectively.

Repository activity is observed via the GitHub REST API repository endpoint, reading stargazers_count and pushed_at. Measured on the same day: mem0ai/mem0 at 68,080, getzep/graphiti at 32,356, letta-ai/letta at 25,292, and langchain-ai/langmem at 1,736. Such figures indicate ecosystem activity only and do not constitute a quality judgment.

B.4 Directions not covered here but worth reading

The following directions are closely related to memory systems but are not developed in this book, because each would need a chapter of its own:

• General context engineering: strategies for compressing, offloading, rebuilding, and validating context. The boundary with memory is lifetime, not data structure.
• Agent observability: standard definitions for traces, metrics, and cost attribution — the upstream input to memory quality measurement.
• Asset registry and version governance: registration, versioning, and release mechanisms for prompts, skills, MCP servers, and agents. Memory entries are governance-analogous to these.
• Behaviour validation and simulation: user and environment simulation before launch, as one means of validating memory strategy changes.

Readers who need to align these directions with memory systems are advised to start from the responsibility-domain division in Chapter 2 and work through item by item, rather than starting from implementation details.