Research and prior art
Message Like Me is a local measurement and profiling layer for message drafting. It is not a model, an autonomous messaging agent, or a claim that a software system represents a person. This review explains the evidence behind that boundary and the neighboring open-source work that informed it.
The cited papers are primary research publications or preprints. Project descriptions link to their official repositories. A paper result is evidence about the task and population it evaluated, not proof that the same result holds for private conversations across the messaging sources a user imports.
Personalization is contextual
PersonaChat found that conditioning a dialogue system on both its assigned profile and information about its interlocutor improved next-utterance prediction. More recent work on linguistic accommodation found that human answers aligned more with a partner's style than LLM answers did, while the LLM answers aligned more closely in semantic content.
These results do not establish a universal method for imitating a person. They support a narrower design inference: a useful messaging profile should separate broadly repeated tendencies from contact-specific and context-specific adjustments. Message Like Me therefore treats incoming messages as response context and only the user's outgoing messages as evidence of the user's prose.
Catch Me If You Can? Not Yet evaluates nuanced individual style in informal communication, a task close to private messaging. Its scope reinforces the same boundary: measured tendencies can guide a draft without establishing a faithful digital copy of its author.
LaMP evaluated personalized classification and generation from user histories and found retrieval-based personalization useful across most of its tasks. Its experiments included term, semantic, and time-aware retrieval. PEARL studied personalized writing assistance and trained a retriever to select historical documents according to their downstream generation value. PEARL also used retrieval quality to identify outputs likely to need revision.
The project inference is selective rather than exhaustive use of history. A small response-context sample with explicit coverage limits exposes less text than placing an entire transcript in an agent prompt. Recency, relationship, and current conversational purpose can all change which past examples apply.
Personalization needs held-out evaluation
ExPerT evaluates personalized long-form generation by comparing evidence-bearing aspects of content and writing style separately. Its reported agreement with human judgment improved over the comparison methods in that study. It does not measure message timing, bubble boundaries, or reply-link behavior.
Can You Make It Sound Like You? studies personalized writing through human review and post-editing. That workflow supports Message Like Me's product boundary: the output is an unsent candidate for the user to inspect and revise, not an autonomous act on the user's behalf.
Münker, Schwager, and Rettinger tested LLM-based imitation of social-network communication and argue that a simulation must be validated for empirical realism in the setting where it was fitted. This supports holding later conversations out of profile creation, drafting from their inbound context without seeing the historical response, and only then comparing the candidate with the reference. A match on surface features alone does not establish semantic equivalence, authorship, or identity.
Digital-agent results are task-bounded
Generative Agents showed that stored experiences, retrieval, reflection, and planning can produce believable agent behavior in a simulated town. Believability in that environment is not the same as fidelity to a real individual.
Generative Agent Simulations of 1,000 People built agents from two-hour interviews with 1,052 participants and evaluated them on surveys, personality measures, and experimental replications. The reported survey result is relative to how consistently participants repeated their own answers two weeks later. It is not evidence that the agents could write private messages like those participants or act on their behalf.
The accompanying official repository does not publish the interview-derived individual agent bank. It describes aggregated access for fixed tasks and reviewed access for individual outputs because of participant privacy. That is useful precedent: an open-source method can remain public while real person-level evidence stays private.
Message Like Me consequently uses the terms style profile, historical tendency, and draft candidate. It does not use a messaging profile to infer beliefs, personality, relationship quality, future behavior, or authority to represent the user.
Private text can remain revealing
Quantifying Memorization Across Neural Language Models found that extractable memorization increased with model capacity, repeated examples, and longer prompting context in the evaluated model families. Beyond Memorization showed that LLMs could infer personal attributes from text even when the task was not extraction of a memorized training example. Removing obvious names is therefore not a complete privacy defense.
When Personalization Misleads found that personalization could steer factual answers toward a user's prior history rather than objective truth in its evaluated settings. The project inference is that meaning, current intent, and factual correctness must outrank style fidelity.
NIST's digital identity risk guidance lists impersonation, privacy loss, and reputational damage among relevant harms. Message Like Me reduces those risks through local storage, bounded exports, pseudonymous ordinary views, explicit provenance, and the absence of message-sending commands. Those controls do not create consent from a conversation partner or make a hosted agent local.
Tempo is relational and descriptive
A 2026 preprint on response times in donated WhatsApp and Instagram chats reported persistent response-speed similarity between chat partners in its sample. This is preliminary evidence from different platforms and cannot set a norm for users of any supported messaging source. It does support comparing tempo within a dyad instead of treating one global latency distribution as a personal rule.
Historical latency is affected by sleep, work, travel, notifications, device availability, urgency, and missing data. Message Like Me reports it as descriptive metadata. It does not tell an agent to wait before returning a draft or portray a historical delay as a preference or promise.
Open-source landscape
| Project | Public scope | Relevant distinction |
|---|---|---|
| OpenSelf | Self-hosted profile and memory system that can automatically reply through WhatsApp, Telegram, and Discord | Message Like Me deliberately ends at an inspectable, unsent draft and does not simulate typing, delay delivery, or operate a messaging account. |
| Second-Me | Locally trained and hosted "AI self" with memory, model alignment, and a network | Message Like Me does not train a model, construct an identity, or join an agent network. |
| Doppelganger | LoRA fine-tuning from chat exports | Its documentation warns that trained models can reproduce private data and other participants' text. Message Like Me keeps analysis in inspectable profiles instead of weights. |
| Write Like Me | Stylometric profiles for several writing registers with held-out verification | It is useful prior art for measured profiles and verification. Message Like Me focuses on dyadic response context, bubble shape, tempo, and reply behavior. |
| imessage-exporter | Broad read-only parsing and export of modern Messages features | It is a valuable compatibility reference. Its GPL-3.0 implementation is not copied into this MIT project. |
| iMessageAnalyzer | Conversation statistics, starters, and successive-message analysis | Its fixed time thresholds are prior art, not universal conversational facts. Message Like Me records its own versioned thresholds with results. |
| iMCP | Sandboxed native access to Messages and the Contacts framework | It demonstrates a possible future native Contacts boundary. The current CLI uses bounded read-only database snapshots. |
| imessage-rag | Local retrieval and question answering over message history | Retrieval of facts from conversations is a different task from measuring the user's outgoing style. |
Explicit limitations
The research above does not establish that:
- one stable profile captures how a person writes to every contact;
- linguistic similarity implies the same intent, judgment, or factual answer;
- a historical response is the response the user would choose now;
- contact frequency, latency, or warmth reveals relationship quality;
- a model-generated draft was authored, approved, or sent by the user;
- pseudonymous identifiers anonymize a corpus against someone with access to the store or installation key;
- local CLI processing controls the data practices of the agent environment that opens a study packet; or
- possession of a conversation database grants permission to publish, fine-tune on, or impersonate its participants.
The defensible claim is smaller: Message Like Me measures selected historical messaging behavior, keeps semantic interpretations tied to bounded evidence, and helps an already-running agent produce an unsent candidate for user review.