# Carry a coding agent's context from one session to the next

> Carry a coding agent's context from one session to the next: 3 companies fully do it, 14 with a limit, 30 carry the tag. Every listing in Agent memory with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/agent-memory.coding-session-memory/
- Markdown: https://www.anchorterminal.com/niches/agent-memory.coding-session-memory/index.md (~2,150 tokens)
- Slim: https://www.anchorterminal.com/niches/agent-memory.coding-session-memory/index.min.md (~380 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/agent-memory.coding-session-memory/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Two or three companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 3 | 14 | 30 | 20 | 8 |

## What counts as done

At the start of a new session the agent gets a summary of the last sessions' decisions and can fetch a cited observation by id in one more call; nothing from another project appears; the hooks do not write files into the repository unasked.

Measured, for our own tests: the summary is under 2,000 tokens.

How agents ask for it: what did we decide about the auth module last time; pick up where we left off; don't re-read the whole repo again

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **AI Memory (Fabio Akita)**: Lifecycle hooks for Claude Code, Codex, Cursor and twenty other harnesses capture sessions, memory_briefing and memory_recent summarise without an LLM, SessionStart sends the handoff, and the wiki lives in the server data directory, not the repository.
- **Basic Memory (Basic Memory LLC)**: The Claude Code and Codex plugins run SessionStart with a bounded context brief and PreCompact checkpoints, capture goes to ~/.basic-memory/inbox not the repository, and bm_read fetches a note by path or permalink.
- **Hindsight (Vectorize, Inc.)**: npx @vectorize-io/hindsight-coding-agents install claude-code adds three hooks, an MCP server and a skill for a per-repo bank, sessions start from a summary of architecture and in-flight work, and uninstall removes precisely what install added.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **claude-mem (Alex Newman)**: Hooks capture every tool use, SessionStart injects the project's recent observations and get_observations fetches cited ids; but the observer queue was lost on restart (issue 4068) and folder CLAUDE.md files, once written everywhere, are now off by default.
- **Cognee (Topoteretes UG (haftungsbeschränkt))**: MCP serves Claude Code, Cursor and Codex with agent-scoped datasets and a code graph, but there are no capture hooks or session summaries; the agent must call remember itself.
- **Context Mode Engine (MKSF LTD)**: Hooks log prompts, edits, commands and errors per project into ~/.context-mode, Claude Code and Codex get a state block of at most 2,048 B at SessionStart, and ctx_resume(seq) returns any row; the block is a log of this machine's sessions, not extracted decisions.
- **EverOS (Evermind AI, Inc.)**: The MCP server reaches Claude Code, Codex and Cursor with an autonomy protocol and a Claude Code plugin saves and recalls sessions, but the agent track is extracted by LLM and the plugin page was not read.
- **Graphicle (KarLex AI, Inc.)**: connect --hooks stores Claude Code and Codex turns as sources and session start gets one line, recall_project takes a token budget and read_project_source continues a cited excerpt; but proposed memories wait for review, and setup writes .claude/settings.local.json and Git hooks into the folder.
- **Honcho (Plastic Labs, Inc.)**: The Claude Code plugin captures through hooks, injects directives, summary and peer card at session start, and redacts secrets; but the memory is about the person across all projects by design, so project isolation is not the model.
- **Hy Memory (Tencent)**: The coding scene stores facts, experiences and a diff chain from tool-call traces isolated by user_id, workspace_id and branch, and source tags separate Cursor, Codex and OpenClaw; capture hooks exist only as Hermes and OpenClaw plugins, none for Claude Code.
- **Mem0 Platform + MCP (Embedchain, Inc. (DBA Mem0))**: The Claude Code plugin captures through hooks, extracts in the background every five exchanges, recalls up to five memories by repo or directory scope and redacts prompts; it needs a Mem0 Platform key, which free accounts must back with their own LLM key after 30 days.
- **MemMachine (MemVerge, Inc.)**: Claude Code connects over MCP with add_memory and search_memory and an OpenClaw plugin exists, but no session capture hooks or summary were found in the pages read.
- **Memori (Memori Labs Inc.)**: A Claude Code skill and memori_recall_summary and memori_compaction restore state across sessions, but the skill fires recall on nearly every turn (listing weakness) and project isolation depends on a process id header.
- **MemPalace (MemPalace (open-source project))**: Stop and PreCompact hooks for Claude Code and Codex tell the AI to file memories and mine imports ~/.claude/projects by wing; but the Stop hook blocks the AI to save, and an upgrade made the palace read as empty until migrate shipped (issue 469).
- **Supermemory API + MCP (Supermemory)**: The Supermemory MCP gives Claude Code and Cursor add_memory, search_memory and get_profile over shared spaces, but no hooks or automatic session capture were found; the agents page covers API integration only.
- **TencentDB Agent Memory (Tencent)**: The proxy records every proxied turn through session/add and injects memory, wiki and code graph into Claude Code, Codex and OpenClaw requests; but session/add is not idempotent, stores all turns, and recall is keyword search with time filters.
- **Tokenrip (Quintel, LLC)**: workspace_load returns bounded context, workspace_sessions lists who worked and each step's why, pins hold up to 20 decisions and tasks carry handoffs; but nothing captures a session automatically.
- **Zep (Zep Software, Inc.)**: The Zep Context plugin and Context MCP Server give Claude Code, Codex and Cursor add_memory and search_graph on the user's own graph, but sign-in runs through Google Workspace or an Enterprise IdP, so a headless agent cannot use it, and nothing captures sessions.

## Couldn't check

The docs were closed to us, or silent. These carry no mark either way.

- **past.dev (Revo Inc.)**: Coding-agent plugins are listed but their pages were not read.

- not this job: Amazon Bedrock AgentCore Memory (Amazon Web Services, Inc. (regional AWS entities by account location)): No hooks or install for Claude Code, Codex or Cursor; events are written by the application through CreateEvent.
- not this job: Egoist AI Passport (Egoist Machines, Inc.): No coding-agent hooks or session capture; the product carries personal memory between assistants.
- not this job: Agent Platform Memory Bank (Google LLC): No hooks or coding-agent integration; memories come from Agent Platform sessions or direct content.
- not this job: Graphiti (Zep Software, Inc.): No hooks or session capture; the MCP server is a general memory graph the agent writes to with add_memory.
- not this job: LangMem (LangChain (the licence file gives no legal form)): No hooks or session capture; a LangGraph library for agents the team builds.

## What stops agents finishing

- observer queue held only in process memory, lost on restart (claude-mem #4068)
- upgrade changes the on-disk format and the store reads as empty (MemPalace #469)
- setup uploads recent prompts to a cloud without asking (claude-mem 15.2.0 to 15.8.3)
- proposed memories wait for a person to approve them (Graphicle)

## Around it

Before: coding-agents.session-end. After: coding-agents.session-start, code-search.repo-context.

## Where this comes from

Weight 20 of 100 in Agent memory: 8 of 36 accounts; the HN 'Stop Claude Code from forgetting' thread (225 comments) and the claude-mem, MemPalace and Basic Memory issues

As data: https://www.anchorterminal.com/api/v1/use-cases/agent-memory.coding-session-memory.json, or the MCP tool get_use_case with id agent-memory.coding-session-memory. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
