Memory

Memory products keep what should survive the context window: user preferences, prior decisions, episode summaries, and extracted facts. Under Agentic, memory is the continuity layer for agents that must not re-interview the company every turn. It is adjacent to RAG (documents) and Vector DBs (embedding stores): many memory systems use vectors, but the product contract is stateful identity over time, not “search this PDF.”

Mechanically you will see memory writes on tool/observation boundaries, consolidation jobs that compress episodes, and retrieval policies that inject only budgeted snippets into the next prompt. Failure modes are silent drift (stale facts treated as current), privacy leaks across tenants, and memory that becomes a second unverified ledger. In OS terms, memory should attach to cases and people with clear authority — not become an ungoverned diary the agent trusts more than the ledger.

Current members:

• mem0 focuses on managed memory layers agents can read/write with less custom plumbing.
• Letta (and similar agent-runtime memory) ties memory to long-running agent state machines.
• Zep emphasizes conversational/temporal memory for assistants.
• LangMem sits closer to framework-native memory helpers in LangChain-shaped stacks.

Prefer dedicated memory when cross-session continuity is the bottleneck; prefer RAG when the missing knowledge is in documents; prefer a graph under Knowledge Graphs when relationships must be typed. Feature maps such as Memory & State Persistence and Learning & Adaptation name the job: durable state that changes only when evidence says so.

Do not confuse chat logs with memory. Logs are append-only observation; memory is curated, queryable, and revisable. If you cannot delete or correct a wrong fact, you do not have memory — you have a liability.

Write policies matter as much as read policies. Memory that accepts every tool observation will drown the next session in noise; memory that never writes will force the agent to re-ask. Prefer explicit schemas (fact, preference, episode) with provenance to free-form “notes.” In multi-tenant installs, memory keys must encode company and case boundaries the same way the ledger does — otherwise one project’s preferences leak into another’s estimate. Consolidate on a schedule; do not wait for context overflow to invent summarization under pressure.

For this research catalog the acceptance test is boring and strict: after wiring the pack, can a second person replay the same scenario and get the same measurable outcome? If not, the page is still a brochure. Document the scenario, the expected number, and the log line that proves it. That habit is what separates a deepened taxonomy cell from a landscape screenshot.