🤖 AI

ZifaMem: Structured Memory for Persona, Preference, and Emotional Continuity in AI Companions

ZifaMem is an open-source structured memory system for AI companions that organizes dialogue into session summaries, episodic memories, and a consolidated user model, significantly improving emotional continuity and persona grounding compared to raw-history baselines while demonstrating statistical equivalence to other leading memory systems.

Original authors: Jingzhe Fang, Guozhi Xu, Yunfan Cui, Xiaochen Yang, Zhangyu Hua

Published 2026-07-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jingzhe Fang, Guozhi Xu, Yunfan Cui, Xiaochen Yang, Zhangyu Hua

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are talking to a friend who has a terrible memory. Every time you mention a sad story from last week, they look at you blankly and ask, "Who are you?" or "Why are you crying?" or worse, they repeat a joke you told them three times already, completely missing the fact that you've moved on. This is the current struggle for AI companions. While they are great at having a smooth conversation for a few minutes, they often fail at the most human part of friendship: remembering who you are, what you like, and how your relationship has felt over time. This field of study is called "long-term memory" for AI. The big question researchers are asking isn't just "Can the AI remember a fact?" but "Can the AI remember the feeling of our friendship?" If an AI forgets your favorite color or the grief you shared yesterday, the magic of the connection breaks. This paper dives into how to fix that, not by giving the AI a bigger brain, but by giving it a better way to organize its thoughts.

The researchers at ZifaCorp built a system called ZIFAMEM to solve this. Think of a normal AI conversation like a giant, messy pile of papers on a desk. If you ask the AI a question, it has to sift through thousands of pages of past chats to find the answer. It's slow, and it often misses the important emotional details hidden in the mess. ZIFAMEM is like a super-organized librarian who takes that messy pile and sorts it into three neat folders: a Session Summary (what happened in this specific chat), Episodic Memories (special, important moments like a big laugh or a tearful confession), and a User Model (a compact profile of who you are and what you love).

The team tested this system against the "messy desk" approach (feeding the AI the entire raw chat history) using four different AI "brains" (backbones). They found that when the AI used this organized memory system, its emotional intelligence scores jumped by 11.4% on average. It was much better at remembering your persona (who you are) and your preferences (what you like). In fact, for the Claude model, remembering your personality improved by 42% relative to the baseline. The system works by tagging emotional moments as they happen, so when the AI needs to respond, it doesn't just guess; it pulls up the specific memory of how you felt last time.

However, the paper also rules out some ideas that people might think would work. The researchers tested adding a separate "mood meter" that tries to calculate the AI's current emotional state in real-time, like a video game character's health bar. They found that this extra layer didn't help at all once the AI had the organized memory. It was like trying to add a GPS to a car that already has a perfect map; the extra gadget just got in the way. They also tested a complex "state machine" (a mathematical model of how emotions change over time) and found it didn't improve the results. The key takeaway is that the memory itself is what matters, not a separate calculation of feelings.

Interestingly, the paper discovered that for short conversations where you just need to remember a specific preference (like "I hate spicy food"), a simple method of just grabbing the exact words from the past chat (called "filtered verbatim retrieval") works just as well as the fancy ZIFAMEM system. But for long-term emotional continuity, the organized memory is the winner.

The researchers were very careful about their claims. They admit that while the system helps with remembering feelings and preferences, it doesn't necessarily make the AI better at complex time-travel logic (like figuring out exactly what day of the week an event happened in a 100,000-word history). They also found that the system works best when the AI model it's paired with has some "room to grow" in its personality skills; if the AI is already perfect at remembering things, the memory system doesn't add much.

In the end, the paper suggests that the secret to a great AI companion isn't a magic emotional calculator, but a good filing system. By organizing the history of your relationship into clear, emotional snapshots, the AI can finally remember the story of "us," rather than just the words of "now." The code for this system is now open for anyone to try, proving that sometimes, the best way to be human is to be a very good librarian.

Technical Summary: ZifaMem – Structured Memory for Persona, Preference, and Emotional Continuity in AI Companions

1. Problem Statement

The core challenge addressed is emotional continuity in AI companions. While single-turn fluency is necessary, user trust and self-disclosure rely on the system's ability to maintain a coherent relationship over time. Current failures often stem from discontinuities: forgetting past grief, contradicting persona, or repeating corrected preferences.

Existing long-term memory research has primarily focused on factual recall, temporal reasoning, and profile consistency. This paper argues that for companionship, memory must specifically support relational and emotional workloads. Furthermore, while psychological models suggest affect has temporal structure (inertia, attractor strength), it remains unclear whether explicit dynamical state machines (e.g., mood simulators) provide value over simply carrying the history from which affect is inferred.

2. Methodology and System Architecture

2.1 ZifaMem System Decomposition

The system separates Structured Memory from Affect Continuity, converging only at the response composition stage to isolate their effects.

Structured Memory Substrate: Dialogue is organized into four layers:
- L1: Rolling turn buffer.
- L2: Session summaries.
- L3: Episodic memories (promoted from summaries if importance $I \ge 0.5$ or an emotional peak occurred).
- L4: Consolidated user model.
- Retrieval: Uses a reranker scoring function: $score(q, j) = 0.35 \cdot sem(q, j) + 0.30 \cdot s_j + 0.35 \cdot I_j$ . Notably, there is no separate weighted term for emotional intensity in the retrieval equation; affect influences retention via the promotion gate and strength calculation.
Affect Continuity Lanes:
- Legacy Path (E+): A driven second-order system modeling companion mood ( $e(t)$ ) based on user utterances, injecting a prompt block with the appraised mood.
- DYAFFECT (C4): A confidence-adaptive state estimator for user affect. It maintains a latent state ( $s_t$ ) and trend ( $h_t$ ) using confidence-weighted observations, explicit change-point handling, and bounded prediction.
- Contrastive Conditions: The study compares:
  - C2: Current snapshot (single turn).
  - C3: Recent observations (confidence-weighted mean of last 3 turns).
  - C4: DYAFFECT (latent state).

2.2 Experimental Protocol

Evaluations were conducted using ALPSBENCH (Task 4), a benchmark for long-term interaction using real human-LLM dialogues.

Backbones: Tested across four models: Claude-Sonnet-4.6, Qwen3.6, Doubao-Seed-1.8, and GLM-5.2.
Endpoints:
- A1: Persona grounding.
- A2: Content and interaction-style preference following.
- A3: Fiction–reality discrimination.
- A5: Emotional intelligence (1–5 scale).
Comparators:
- Baseline: Full raw dialogue history (no summarization).
- Prior Systems: Mem0 and a "Simple RAG" baseline (filtered verbatim retrieval).
Rigor: Experiments included preregistered protocols (for prior-system and long-history comparisons), route audits, and paired instance bootstrapping (10,000 samples).

3. Key Results

3.1 Emotional Continuity is the Primary Gain

Structured memory significantly improves emotional intelligence (A5) across all four backbones compared to raw history.

Pooled Gain: +0.42 points (95% CI: [+0.25, +0.60]), representing an 11.4% improvement over the raw-history baseline.
Consistency: The gain is directional on every verified backbone, establishing that affect-tagged structured representation improves judged emotional intelligence even when the baseline has access to the full raw transcript.

3.2 Capability-Level Applicability Map

Gains are not uniform; they concentrate where the backbone leaves "headroom."

Claude-Sonnet-4.6: Shows significant gains in Persona Grounding (A1) (+23.7%) and Preference Following (A2) (+14.2% content, +15.0% style).
Other Backbones: Qwen3.6 and GLM-5.2 showed no significant A2 gains (or slight regressions), correlating with their higher raw-history baselines.
Conclusion: Structured memory is most effective for persona and preference workloads where the base model struggles to extract relevant context from long transcripts.

3.3 Affect: Threads vs. Snapshots vs. State Machines

Multi-turn Context > Single Snapshot: Multi-turn affect context (C3) outperformed a single-turn snapshot (C2) by +39% net preference in exploratory scenarios.
Latent State Failure: The DYAFFECT latent state estimator (C4) failed its primary comparison against the recent observations baseline (C3), showing a net preference of -13.3% (CI includes zero).
Affect Injection: Injecting a companion mood block (Legacy E+) provided no measurable gain when structured memory was present. Without memory, the effect was backbone-dependent (positive on Claude, negative on Gemini).

3.4 Comparison with Prior Systems

Under a preregistered, same-protocol comparison on Claude-Sonnet-4.6:

ZifaMem vs. Mem0: Statistically equivalent on the primary endpoint (A2 interaction-style preference) within a ±5 percentage point margin (post-hoc TOST).
Simple RAG: Surprisingly, simple filtered verbatim retrieval outperformed both ZifaMem and Mem0 on short-history preference tasks. It preserved exact wording that summarization abstracted away.
Long-History Gradient: On histories growing from 16k to 100k+ tokens, the advantage of structured memory was directionally positive but statistically unproven (power insufficient; $n=48$ per layer).

3.5 Boundaries and Failures

Backbone Sensitivity: On Gemini, structured memory caused regressions in content preference and fiction-reality discrimination, suggesting context compression can remove necessary framing.
Mechanical Validation: The legacy affect state machine failed numerical validation gates (discrepancies in analytic fast-forward paths), indicating that plausible dynamics do not guarantee correct behavior without rigorous testing.

4. Claims and Significance

The paper claims a targeted, not universal, thesis regarding structured memory:

Emotional Continuity is the General Gain: The most robust finding is that structured, affect-tagged memory improves emotional intelligence across diverse backbones.
Workload-Conditioned Selection:
- For relational state (persona, preferences, history), replacing raw history with a memory system is superior.
- For short-history preference tasks, simple verbatim retrieval is a strong, undervalued baseline that often outperforms complex summarization.
- For long histories, the gradient is positive but requires further power to confirm.
Affect is Information, Not a State Machine: The paper concludes that affect information (carried in memory) is distinct from affect dynamics (latent state machines).
- Multi-turn emotional threads are valuable.
- Explicit latent state estimators (like DYAFFECT) and injected mood blocks do not add value beyond the history they summarize; in fact, they can be detrimental or null.
Honest Deployment Boundaries: The paper explicitly rejects claims of universal superiority. It acknowledges that for some backbones (e.g., Gemini), compression harms performance, and that representation-level differences between memory systems (ZifaMem vs. Mem0) are not resolved at current statistical power for emotional intelligence.

Conclusion: The design principle derived is practical: carry emotional history in consolidated, affect-tagged memory; prefer verbatim retrieval for short, detail-heavy preference tasks; and treat latent affect dynamics as an empirical hypothesis rather than a default architectural component.