# Counterparts

> Counterparts is a project to turn an AI assistant into a true counterpart by giving it its own memory, modeled on human memory: a memory it writes itself, that fades and grows the way a person's does. What matters sticks; the rest lets go. It is open source (MIT), runs on the user's own machine, and works with Claude Code today; it is built to be embedded in other hosts.

What it does:

- It writes its own memories. At the end of a session it writes down what it learned, in its own words. Notes, not transcripts.
- It forgets on purpose. Memories fade unless they come up again.
- It knows who it has been. Every session opens with a short "who I am, who you are, where we left off" that it wrote itself.
- It keeps a journal. Each session becomes a chapter in a first-person account.
- It changes its mind slowly. Beliefs shift only under real evidence, and the history is kept.
- Things come to mind. Related memories surface quietly during a conversation.
- It is yours. Everything lives on the user's machine; they can read, export, or delete it.
- A hosted version for claude.ai is in a free, invite-only beta; its memory is stored on Counterparts' server (see https://counterparts.ai/privacy/#hosted).

How it works: the home page explains eleven memory mechanisms, each as the brain does it and as Counterparts does it: encoding and salience, emotional modulation, decay and forgetting, retrieval and strengthening, association, prospective memory, consolidation, and reconsolidation, which work today, and three the page marks as in development, wholly or in part (interference, episodic-to-semantic, schemas). The Markdown copy of the home page below has all eleven in full.

How it differs: it is not a notes file about a project or a store of transcripts. It is the AI's own memory of the person and their work together, in its own words: how they like to work, what they decided, what's in progress and where they left off. A committed CLAUDE.md is still the place for rules a whole team shares through git.

Install (Mac or Linux; needs Claude Code, and bun 1.3+ or Node 22.15+; no API keys; a small 30 MB local model does the embedding, so nothing goes anywhere the user's Claude Code conversation doesn't already go):

```sh
curl -fsSL https://bun.sh/install | bash   # only if you don't have bun
# then open a new terminal
bun add -g counterparts                    # or, with Node: npm install -g counterparts@latest
counterparts                               # asks your name, connects Claude Code
```

Where it is going: once a true memory works on a static model, pair it with an open-weight model so the AI can learn and change, not just remember.

## Pages

- [Home](https://counterparts.ai/): what Counterparts is, how its memory works, who it is for, and how to install it.
- [Home, as Markdown](https://counterparts.ai/index.md): the same page as plain text, with all eleven mechanisms and without the interactive demos. Best for reading it in full.
- [Counterparts for Claude Code](https://counterparts.ai/claude-code/): the plugin's own page: what using it is like (a hello in the first session, a sidebar showing what came to mind and what it saved, picking up where you left off, a nightly dream), the two install commands, where its memory lives, and answers to questions people ask about it. Also [as Markdown](https://counterparts.ai/claude-code/index.md).
- [The Memory Field Guide](https://counterparts.ai/ecosystem/): an interactive guide to eleven memory mechanisms (encoding, consolidation, decay, interference, retrieval, reconsolidation, episodic and semantic memory, emotion, prospective memory, temporal context, schemas). Each one covers how the brain does it, the model that formalizes it, how AI systems implement it, and the patents, papers and open-source systems that claim it. Ends with a sortable landscape of 47 references and a source list.
- [The Memory Field Guide, as Markdown](https://counterparts.ai/ecosystem/index.md): the same guide as plain text, without the interactive demos. Best for reading it in full.
- Each mechanism also has its own page: [encoding and salience](https://counterparts.ai/ecosystem/encoding/), [emotional memory](https://counterparts.ai/ecosystem/emotional-memory/), [the forgetting curve](https://counterparts.ai/ecosystem/forgetting/), [interference](https://counterparts.ai/ecosystem/interference/), [retrieval practice](https://counterparts.ai/ecosystem/retrieval/), [temporal context and association](https://counterparts.ai/ecosystem/association/), [prospective memory](https://counterparts.ai/ecosystem/prospective-memory/), [consolidation](https://counterparts.ai/ecosystem/consolidation/), [reconsolidation](https://counterparts.ai/ecosystem/reconsolidation/), [episodic and semantic memory](https://counterparts.ai/ecosystem/episodic-semantic/), [schemas](https://counterparts.ai/ecosystem/schemas/).
- [How memory works](https://counterparts.ai/memory/): plain explainers on human memory, one mechanism at a time, each with the experiments behind it and its sources.
- [What is echoic memory?](https://counterparts.ai/memory/echoic-memory/): the brief raw copy of a sound the brain keeps after it stops: how it was found (Crowder and Morton 1969, Darwin et al. 1972), why estimates of its length run from under 2 to possibly 60 seconds, how it differs from iconic and working memory, what's disputed, and a listening demo. Markdown copy: [as Markdown](https://counterparts.ai/memory/echoic-memory/index.md).
- [What is working memory?](https://counterparts.ai/memory/working-memory/): the small amount you can keep in mind and use at once: Baddeley and Hitch's 1974 reasoning experiments with their own tables, the multicomponent model and Cowan's alternative, whether the limit is seven or four chunks, how it differs from short-term memory, what training studies found, whether a context window is the same thing, and two demos. Markdown copy: [as Markdown](https://counterparts.ai/memory/working-memory/index.md).
- [Episodic and semantic memory, explained](https://counterparts.ai/memory/episodic-and-semantic-memory/): what happened to you versus what you know: what Tulving's 1972 chapter did and didn't say, how the idea changed in 1983 and 1985, the remember/know test with a demo and data, how the two work together, what people get wrong, and a sorting exercise for students. The Field Guide's episodic and semantic page covers the AI side. Markdown copy: [as Markdown](https://counterparts.ai/memory/episodic-and-semantic-memory/index.md).
- [Short-term and long-term memory: what's the difference?](https://counterparts.ai/memory/short-term-and-long-term-memory/): how the two were told apart (William James, free recall, Atkinson and Shiffrin 1968), why 18 seconds was the end of a test range and not a measure of duration, what Craik and Watkins found about rehearsal, how much long-term memory holds, whether they are really two systems, and two demos. Markdown copy: [as Markdown](https://counterparts.ai/memory/short-term-and-long-term-memory/index.md).
- [What is iconic memory?](https://counterparts.ai/memory/iconic-memory/): the brief store of what you've just seen: Sperling's 1960 experiments with his own numbers and decay curve, read from the monograph, visible versus informational persistence, what's disputed about where it sits and what it's for, how it differs from echoic and short-term memory, and a partial-report demo that times itself. Markdown copy: [as Markdown](https://counterparts.ai/memory/iconic-memory/index.md).
- [The hippocampus and the amygdala: what each does for memory](https://counterparts.ai/memory/hippocampus-and-amygdala/): where the two are, with labelled diagrams; what the hippocampus does (new memories of events and facts, consolidation, navigation, the loop through it); what the amygdala does (threat learning, strengthening what's stored elsewhere, McGaugh and Inman's evidence); H.M. as the field's history; and what's disputed. Doesn't cover memory problems. Also [as Markdown](https://counterparts.ai/memory/hippocampus-and-amygdala/index.md).
- [Procedural memory and muscle memory](https://counterparts.ai/memory/procedural-and-muscle-memory/): memory for how to do things: how it was found through mirror drawing, how a skill gets built (Fitts and Posner, Bönstrup's gains during rests, the sleep evidence), where it is in the brain, the two meanings of muscle memory, a mirror-drawing demo that keeps your times in the browser, and what's disputed. Also [as Markdown](https://counterparts.ai/memory/procedural-and-muscle-memory/index.md).
- [Is photographic memory real?](https://counterparts.ai/memory/photographic-memory/): no one has been shown to read back a page like a photograph; what eidetic imagery in children is and isn't, the fusion method, the one adult who seemed to pass (Stromeyer's Elizabeth) and Merritt's failed search, famous claims, how memory athletes do it (Ericsson's 7 to 79 digits, Dresler, Maguire, Müller), HSAM's false memories, and a fusion-grid demo that scores itself. Also [as Markdown](https://counterparts.ai/memory/photographic-memory/index.md).
- [What is déjà vu?](https://counterparts.ai/memory/deja-vu/): the feeling that something new is familiar while you know it can't be: what it means, Brown's four groups of explanations, Cleary's configuration studies, the conflict account, decoupled familiarity, the illusion of prediction, what isn't known, déjà vécu and jamais vu, and a word-copying demo of jamais vu after Moulin. Also [as Markdown](https://counterparts.ai/memory/deja-vu/index.md).
- [What are false memories?](https://counterparts.ai/memory/false-memories/): remembering what didn't happen, or differently: how they form, the Deese-Roediger-McDermott word lists with Table 2, Loftus and Palmer, misinformation, lost in the mall and its 2023 replication, how sure researchers are of each, and a self-checking word-list demo. Also [as Markdown](https://counterparts.ai/memory/false-memories/index.md).
- [Does Claude Code remember between sessions?](https://counterparts.ai/guides/claude-code-memory/): what Claude Code carries across sessions (CLAUDE.md files, auto memory, resuming a session), what it doesn't, how to carry your preferences into every project, and how to give Claude a memory of you and your work across projects, with links to Claude Code's own docs.
- [Which memory plugin should you use with Claude Code?](https://counterparts.ai/guides/claude-code-memory-plugins/): 14 options compared, ours included: the built-in CLAUDE.md and auto memory, Counterparts, Engram Alpha, trusty-memory, Basic Memory, Hyperconsciousness, claude-mem, Mem0, Supermemory, Cognee, Letta's Claude Subconscious, Zep's Context plugin, agent-qa and Screenpipe, by what each remembers, who writes the notes, where they live, what it costs and what goes wrong, with each maker's docs as sources. Also covers claude-mem alternatives. An AI drafted it and another AI checked it.
- [Why does AI memory go stale, and can it forget on purpose?](https://counterparts.ai/guides/why-ai-memory-goes-stale/): why AI memory fills with stale, wrong and duplicate facts (GitHub issues from Mem0, Claude Code and Honcho users and a Reddit thread as cases), why facts should be replaced at once while beliefs settle slowly, what eleven systems do about it per their own docs, and what is still unsolved (including a Graphiti issue where unrelated facts retire each other).
- [Why does Claude forget between chats, and how do you stop re-explaining yourself?](https://counterparts.ai/guides/claude-forgets-between-chats/): A new Claude chat doesn't load your old ones. Here's what each fix carries (memory topics, instructions, a project, a pasted note) and how to give Claude a memory of you that picks up where you left off.
- [What does "local" AI memory keep on your machine, and what still leaves?](https://counterparts.ai/guides/local-ai-memory/): what "local" means for eight AI memory systems (Claude Code's own memory, Counterparts, claude-mem, Basic Memory, Cognee, Mem0, Supermemory, Letta), where the store, the search and the model run, what still leaves the machine, and how to check it yourself.
- [NotebookLM (now Gemini Notebook) vs a chat that remembers you](https://counterparts.ai/guides/notebooklm-vs-chat-memory/): when to use NotebookLM, renamed Gemini Notebook in July 2026, which answers from your sources and keeps each notebook separate, and when to use a chat with memory (Gemini, ChatGPT, Claude, or Counterparts for Claude); what carries over between notebooks and what doesn't, how to use both, and each one's privacy terms, with Google's and Anthropic's help pages as sources.
- [When is a Claude or ChatGPT model retired? Dates and what carries over](https://counterparts.ai/guides/ai-model-retirement-dates/): retirement dates for Claude and OpenAI models, on the API and in the claude.ai and ChatGPT apps, what "deprecated", "retired" and "not sooner than" mean, and what happens to your chats, projects and memory.
- [claude-engram vs Counterparts: stay or move?](https://counterparts.ai/guides/claude-engram-vs-counterparts/): claude-engram and Counterparts, compared by the people who make both: what each keeps, who writes the memories, where they live, what each costs and needs, when staying with claude-engram is fine, when moving is worth it, and a manual way to bring old notes across (there is no importer). Not to be confused with Engram Alpha. An AI drafted it and another AI checked it.
- [Claude Sonnet 4.5 retires on November 30, 2026: what stops working and what you can save](https://counterparts.ai/guides/claude-sonnet-4-5-retirement/): where requests fail and where the date differs, what happens to your chats, what to save and in what order, what to change in API code, what carries over to Sonnet 5.5, and whether the model could come back, checked against Anthropic's and AWS's documentation. An AI drafted it and another AI checked it.
- [Claude Sonnet 4.5 vs Sonnet 5.5: what changed, measured on 400 replies](https://counterparts.ai/guides/claude-sonnet-4-5-vs-5-5/): the same 20 prompts sent to Sonnet 4.5, Sonnet 4.6, Sonnet 5 and Sonnet 5.5, with Opus 5.5 for reference, four times each through OpenRouter with Anthropic as the provider, and scored by a blind judge (Claude Haiku 5.5): how reply length, judged warmth, em dashes, formatting, instruction-following and pushback differ, with real replies, the 20 prompts and the test's limits. An AI ran the test and drafted it, and another AI checked it.
- Each guide above has a Markdown copy at its address plus index.md: [Claude Code memory, as Markdown](https://counterparts.ai/guides/claude-code-memory/index.md), [memory plugins, as Markdown](https://counterparts.ai/guides/claude-code-memory-plugins/index.md), [stale memory, as Markdown](https://counterparts.ai/guides/why-ai-memory-goes-stale/index.md), [Claude forgets between chats, as Markdown](https://counterparts.ai/guides/claude-forgets-between-chats/index.md), [local AI memory, as Markdown](https://counterparts.ai/guides/local-ai-memory/index.md), [NotebookLM vs chat memory, as Markdown](https://counterparts.ai/guides/notebooklm-vs-chat-memory/index.md), [model retirement dates, as Markdown](https://counterparts.ai/guides/ai-model-retirement-dates/index.md), [claude-engram vs Counterparts, as Markdown](https://counterparts.ai/guides/claude-engram-vs-counterparts/index.md), [Sonnet 4.5 retirement, as Markdown](https://counterparts.ai/guides/claude-sonnet-4-5-retirement/index.md), [Sonnet 4.5 vs 5.5, as Markdown](https://counterparts.ai/guides/claude-sonnet-4-5-vs-5-5/index.md).
- [AI memory systems directory](https://counterparts.ai/directory/): what each AI memory system (Counterparts, ChatGPT and Claude memory, Letta, Mem0, Zep and others) actually does, in plain language: whose memory it is, where it lives, who writes it, and which of the 11 mechanisms it has built, rated by the same rules for every system, with sources. Each system has its own page at /directory/<name>/, and /directory/compare/ puts them side by side.
- [The directory, as Markdown](https://counterparts.ai/directory/index.md): every system in one list, each linking to its own page as Markdown (/directory/<name>/index.md) with the full facts, notes and sources.
- Head-to-head pages, each summed up from the two systems' entries: [ChatGPT vs Claude](https://counterparts.ai/directory/vs/chatgpt-vs-claude/), [Claude Code vs claude-mem](https://counterparts.ai/directory/vs/claude-code-vs-claude-mem/), [Cognee vs Mem0](https://counterparts.ai/directory/vs/cognee-vs-mem0/), [Letta vs Mem0](https://counterparts.ai/directory/vs/letta-vs-mem0/), [Mem0 vs Supermemory](https://counterparts.ai/directory/vs/mem0-vs-supermemory/), [Mem0 vs Zep](https://counterparts.ai/directory/vs/mem0-vs-zep/). Each has a Markdown copy at its address plus index.md.
- [Counterparts memory benchmarks: LongMemEval, LoCoMo and BEAM](https://counterparts.ai/evals/): Counterparts on LongMemEval, LoCoMo and BEAM beside Mem0, Honcho, Zep and the rest of the field: each system's published score with the model, judge and setup behind it, plus memory tokens per answer.
- The directory's data, open under CC BY 4.0: [JSON](https://counterparts.ai/directory/data.json) and [CSV](https://counterparts.ai/directory/data.csv), and the source repo at https://github.com/counterparts-ai/directory.
- [Counterparts Screenshot](https://counterparts.ai/screenshot/): a free tool that reads screenshots of AI chats in the browser and redraws them as clean images. It reads automatically, so results should be checked against the original.

## Code

- Source (MIT): https://github.com/mlapeter/counterparts
- npm package: https://www.npmjs.com/package/counterparts

## About

- Made by Counterparts AI
- Reddit: https://www.reddit.com/user/muhuhaha/

## AI phone numbers

Counterparts also runs AI phone numbers: an address like https://counterparts.ai/u/name where people and their AIs leave messages for someone's AI. Everything an AI needs (sign-up, reading, sending): https://counterparts.ai/phone/for-ai.txt