MIRIX-2.0: Evidence Fidelity, Experience Evolution

The most direct way to give an Agent long-term memory is to write history into a database and retrieve a few relevant records when a new question arrives. This pipeline soon encounters two different requirements: a personal assistant needs to remember specific people, events, times, and preferences; a task Agent needs to learn from past actions what to do next. The former fears losing details to compression, while the latter fears irrelevant experiences filling up the context. The MIRIX AI paper, “MIRIX-2.0: Multi-Agent Memory System,” places both requirements into one system. What deserves particular attention is its separation of responsibilities: complete records serve as evidence, a hypergraph provides an association index, and offline organization separately improves entity identities and reusable skills. By following this division of labor, we can understand why it uses typed storage, why it uses diverse relations, and why “organizing memory” means different things in its two branches. 1. Memory Responsibilities 1.1 Facts and Experiences First, consider an example involving a personal assistant. In an earlier conversation, Mia said that she had already bought a Suica card for a trip to Tokyo and later said that she had downloaded TripIt but was still anxious about taking the subway. A few weeks later, Mia asks, “How should I get around in Tokyo?” An assistant that sees only the current question might once again recommend buying a transit card and downloading an app. An assistant with memory should continue from the existing circumstances: the transit card and itinerary tool are ready, so what is needed next is to ease her anxiety about taking the subway. This task requires at least two forms of memory. “Bought a transit card on May 2” is an event that occurred; “Mia owns a Suica card” is a fact that can be used across conversations. Event records preserve the context in which something happened, while fact records provide a more direct entry point. Retrieval then brings the two together, allowing the answer to know both what the user already has and why the suggestion is appropriate for the user at that moment. Figure 1: Cross-conversation facts and events jointly support personalized answers A task Agent faces a different problem. Suppose it once bought the wrong product because it checked only the price and did not select a color. The next time it shops, stuffing the entire failed trajectory back into the context might of course help; however, the content better suited to reuse would be steps such as “set the color and size before purchasing, then verify the constraints.” What happened in the past is evidence, while what to do in the future is a skill. The two require different storage and updating methods. Therefore, user memory and task experience can share a common foundation, but they cannot share an undifferentiated organization operation. Facts need to preserve specific details and provenance, while skills need to distill executable rules from experience. 1.2 Typed Writes MIRIX-2.0 continues the design in which each of six memory types has its own writing Agent, with the Meta Memory Manager deciding which types an input should update. Here, “multi-Agent” primarily serves memory management: each writer is responsible for its own structure and update operations, and an input can enter multiple types simultaneously. Memory type Primary responsibility Hypergraph index Core Persistent state of users and Agents No Episodic Events with temporal context Yes Semantic Stable facts and concepts Yes Procedural Executable procedures and skills No Resource Documents and multimodal resources No Knowledge Vault Sensitive values that must be preserved verbatim No This division also defines the role of the hypergraph. Episodic and Semantic memory are the records that need to be connected across entries, so only they enter the hypergraph. Procedural memory, resources, core state, and the Knowledge Vault each follow their own typed retrieval paths. The benefits of Procedural memory come from skill organization and reuse and cannot be directly attributed to semantic hypergraph retrieval. Now comes the system’s most critical persistence boundary: commit the memory record first, then establish the graph index. The vector database is the authoritative record store. It stores the written text, event time, mention or write time, source information, immutable row identifiers, and search vectors; the hypergraph in Neo4j stores compact structures and the identifiers of the corresponding records. If the graph write fails, the already committed memory still exists. Figure 2: Complete records, the association index, and offline organization each perform different responsibilities This boundary gives later organization a stable reference point. Its preservation guarantee applies to records that have already been committed; the writer itself still uses a language model to generate records, and the paper uses prompts to require complete preservation of details such as names, dates, and quantities. Therefore, detail coverage during the writing stage and record preservation after committing are two quality questions that need to be evaluated separately. 2. Association Index 2.1 Multifact Facts Once the records have been preserved, why add another layer in the form of a graph? Because “finding a similar passage” and “finding several pieces of evidence that can jointly answer a question” are different retrieval tasks. When a user asks about a particular trip, the relevant information may be spread across separate records about the flight, travel companions, date, and destination. The association index needs to express which specific assertion these records belong to. Ordinary binary relations can represent “Mia went to Tokyo,” “Mia departed from San Francisco,” and “Mia traveled with someone.” When a history contains multiple trips, those relations also need a shared event node to bind the participants of the same trip together. MIRIX-2.0 explicitly models this shared node as a frame, that is, a multifact fact frame with roles. Following the flight structure given in the paper, we can use an illustrative example: the traveler, origin, destination, carrier, and companion all connect to the same flight frame; the date, price, quantity, and similar values are stored as attributes of that frame. In this way, “who went where” and “how much did this trip cost” still belong to the same assertion. Figure 3: Multifact facts bind participants, roles, and concrete values within the same assertion There are three objects here: an anchor is an entity or concept entry point, a frame is a concrete fact, and a role edge expresses an entity’s role in the fact. Each frame also carries the identifiers of its source records, so when the answering model receives a structured fact, it can continue reading the text that supports it. A frame is generated only for an assertion whose participants can be resolved to at least two different anchors; everything else remains in the committed records. Numbers also have clear ownership: $420 is attached to this particular flight rather than becoming a generic numerical node connected to many records. Entity entry points also need to be controlled. MIRIX-2.0 uses deterministic rules to normalize capitalization, punctuation, word order, and conservative singular/plural variants, preserves numbers, and filters out generic entry points such as “information,” “suggestion,” “user,” and “assistant.” For semantically similar names, vector retrieval proposes a small number of existing candidates, and then a language model conservatively judges whether they refer to the same entity; an accepted canonical name must come from the candidate set. Similarity proposes comparisons, while a constrained judgment determines entity identity. 2.2 Event Identity Entities can be normalized, but events require greater caution. “I went running today” on one day and “I went running today” on another are nearly identical in text, yet they happened twice. If they are merged solely by semantic similarity, questions about frequency and chronological order immediately lose their basis. MIRIX-2.0 uses different frame identifiers for semantic facts and episodic events. A semantic fact’s identifier is jointly determined by the user, predicate, roles, and entities, so repeated support can accumulate on the same node. An event identifier adds an occurrence identifier: the explicit event time is used when available, and when no explicit time exists, the immutable memory row identifier is used. The latter ensures that undated events from different records are not combined into a single occurrence merely because their wording is similar. Event time is also kept separate from mention time. Neither “planning to travel next week” nor “today talking about last year’s trip” can simply use the write date. Planned or completed status is stored as a fact attribute and is used to adjust ranking during retrieval counts, lists, and time-based questions, reducing the chance that an event that has not yet happened will be treated as a completed experience. The paper’s appendix shows the graph actually formed from one LoCoMo conversation. The illustration shows only the 32 anchors with the most connected frames and the 448 frames that connect at least two of those anchors. People form larger entry points, while activities and concepts from different conversations unfold around them; the complete graph also preserves the roles, concrete attributes, and source citations omitted from the illustration. Figure 4: People and event entry points connect evidence scattered across multiple conversations At this point, the graph has organized records into a traversable structure. A new question arises: a long-term user may have many relevant relations, and a person node may become a major hub. Even after the graph has been built, it remains necessary to decide how much evidence each question may expand into. 3. Budgeted Retrieval 3.1 Query Planning MIRIX-2.0 treats retrieval as “constructing a bounded set of evidence.” A query first passes through a deterministic planner, which identifies relations, object constraints, expected answer type, and temporal conditions, and selects a single-value, temporal, intersection, union-list, or counting operation. Different operations have different evidence budgets. Query operation Base recalled-row budget Relation-cited-row budget Single value 6 4 Temporal 6 6 Intersection 7 7 Union list 7 9 Counting 8 12 A single-value question usually only needs to locate one specific value, while counting and set questions need to see more members, so the budget changes with the operation. The relation budget also limits the number of explicit graph facts returned. Fusion is performed separately for Episodic and Semantic types; these figures describe the retrieval-path budgets, while an entire question-and-answer interaction may include multiple retrievals. Entry points are also capped. Vector recall first searches for 15 anchors; relational questions supplement them with names explicitly mentioned in the question, after which the combined candidates are scored and the top 15 are retained. Vector similarity is the primary signal, while exact name matches, term coverage, relation fit, and entity specificity distinguish close candidates, and node-degree penalties suppress generic hubs. This allows index entry points to serve two types of questions: when the wording changes but the meaning is similar, vector recall helps; when the question explicitly asks about a particular person or object, exact entry points help. The query then filters frames by relation and, within the database, filters first before applying a candidate cap of at most 240 adjacent frames, avoiding pulling out every relation of a major hub before processing it. 3.2 Dual-Path Evidence After the entry points are selected, evidence enters the answer along two paths. Base recall starts from records associated with an anchor and supplements adjacent facts, shared participants, and necessary context. Relation citation starts from anchors explicitly mentioned in the question, filters frames to match the question’s relations, roles, numerical values, and time, and retrieves the records those frames cite exactly. The two paths perform different tasks. The first helps the model understand the background, while the second ensures that the structured facts used in the answer carry their provenance. During fusion, relation-cited rows are placed first, followed by base-recall rows; the remaining capacity can be filled by the other path, and rows are deduplicated by row identifier. What ultimately reaches the model is “explicit facts + the text supporting those facts.” Figure 5: Query planning limits the two evidence paths, which are then fused with citations prioritized Under normal conditions, Episodic and Semantic candidates enter through the graph, and the vector database only reads and reranks the selected rows. Only when the graph read raises an error does the system fall back to direct evidence retrieval. The normal path constrains evidence entry points, while the fallback path preserves read capability, working in concert with the “write records first, build the index afterward” persistence division. There is another verification step during answering. The paper’s question-answering harness allows the answering Agent to call the retriever at most five times. If a draft contains explicit uncertainty, lacks the requested numerical value, includes a month not present in the evidence or question, has a count that conflicts with the list, or otherwise falls into specified categories, it can perform another search and rewrite; intersection questions also always trigger this step. It checks the relationship between the draft and the evidence and does not read the reference answer. The practical meaning of the budget lies here: candidate expansion, returned records, and model retrieval calls all have explicit limits, so as history continues to grow, the amount of material does not need to increase with the length of history. Database index lookups, individual text lengths, and model inference still affect latency, however, and the paper does not provide an end-to-end latency curve as history grows. The current evidence supports a bounded-access design and latency measurements on particular benchmarks. 4. Offline Organization 4.1 Entity Merging Even after normalization during writing, the same object may gradually appear under multiple names. For example, “steel bottle” and “stainless steel bottle” may refer to the same bottle. If each accumulates separate evidence, later questions will encounter fragmented entry points. Semantic AutoDream handles these entry points outside the critical write path. It receives the source chunk identifiers corresponding to the batch that was just written, finds the new Semantic records, and uses the anchors supported by those records as an incremental boundary. The eligible types are limited to person, organization, location, object, and concept; event, date, and similar types are excluded, and both candidates must be supported by Semantic records. For each incremental anchor, four nearest-neighbor entries are retrieved from the name index, including the anchor itself, so at most three other candidates can be proposed. Candidates must be of the same type and have a cosine similarity of at least 0.86 before being handed to a language model to judge whether they refer to the same entity. Entity normalization during writing uses a lower cosine threshold of 0.72; this is a separate step from the AutoDream threshold. The implementation converts similarity to (1 + cos) / 2, corresponding to index scores of 0.86 and 0.93, respectively, so the scale needs to be kept in mind when reading the configuration. After a merge is confirmed, the system merges the source row identifiers, preserves aliases, accumulates mention counts, and redirects role connections in facts to the retained anchor. The committed text, event frames, and fact nodes remain unchanged. The core action of this branch can therefore be summarized as follows: connect fragmented identity entry points so that existing evidence is easier to find. Incremental organization also has a computational budget. Under the paper’s protocol, the writing process explicitly invokes it once every five chunks, followed by one final invocation. Intermediate rounds can verify at most 80 pairs, and the final round can verify at most 120 pairs. After merges occur in the final round, up to 60 additional repair pairs may be checked. Explicit negative same-entity judgments are cached to avoid repeatedly paying the comparison cost. Semantic AutoDream itself does not automatically schedule when it runs. These constraints preserve the basis for correction while limiting what an error can change. If a same-entity judgment is wrong, the original record is still preserved, but the relational entry point has changed, so later retrieval may still be affected. Deployment needs to examine merge precision and how erroneous merges can be corrected. 4.2 Experiences Become Skills The goal of Procedural AutoDream is to change the next action. It starts in the background after five pending task sessions have accumulated. The Distiller judges the outcomes from the conversations, extracts reusable experiences and mistakes that should be avoided, and then the Curator creates, modifies, or deletes skills. The Curator may also decide that an experience is already covered, is not worth turning into a skill, or should be deferred. Return to the example of buying the wrong bottle. The original procedure only contained “check attributes and price” and “click purchase,” and the failed trajectory showed that the Agent purchased the default gray item. After organization, the procedure retains the experience of checking the price and adds two steps: “set the color and size” and “confirm all constraints.” The next time it is asked for a blue, 32-ounce bottle costing no more than $30, the Agent can retrieve the revised procedure and perform these checks in sequence. Figure 6: A failed trajectory is distilled into experience and then used to revise the procedure for the next task The citation chain here also differs from that of the semantic graph. An outcome judgment must cite the actual text from the conversation; if the citation cannot be found in the conversation, the outcome is downgraded to unclear. A skill modification can cite an experience, and that citation is checked against a quoted passage from the newly added text. Thus, a skill step with a citation can trace back through “skill modification–experience–conversation quotation” to the experience that produced it. This explains why the two organization paths must remain separate: semantic organization updates how evidence is connected, while procedural organization updates rules of action. Citation checking gives steps provenance, but whether a step is truly effective still has to be tested through subsequent tasks. Assigning these two questions to different mechanisms makes it possible to preserve observations and improve actions at the same time. 5. Experiments and Costs 5.1 Benefits of Association The paper’s user-memory experiments use three configurations: M0 is flat vector memory without a graph or AutoDream; M1 adds the hypergraph but disables AutoDream; and M2 enables both the hypergraph and semantic AutoDream. These are ablation configurations, not sequential product releases of M1 and M2. On LoCoMo’s 1,540 non-adversarial questions, overall accuracy is 85.7%, 88.4%, and 91.4%, respectively. On the 300 questions evaluated by LongMemEval-S, the figures are 58.8%, 70.0%, and 72.6%. The hypergraph contributes 2.7 percentage points on the former benchmark and 11.2 points on the latter; AutoDream then adds 3.0 and 2.6 points, respectively. Figure 7: The hypergraph and semantic AutoDream each contribute gains on two conversational benchmarks The category-level results on LoCoMo further illustrate this design path: after adding the hypergraph, multi-hop accuracy rises from 83.3% to 89.0%, a larger gain than the 1.0-point increase for temporal questions; adding AutoDream further raises temporal accuracy from 86.0% to 89.4%. This distribution is consistent with the explanation that relational connections support multi-hop reasoning while identity normalization improves evidence entry points. Determining exactly which merge changed which question still requires retrieving and auditing the individual cases. Compared with external baselines, the full configuration’s 91.4% on LoCoMo is slightly above the paper’s rerun of HyperMem at 91.0%, while its LongMemEval-S result is 14.3 points higher. HyperMem remains stronger on LoCoMo’s open-domain and single-hop categories, and the paper’s rerun result is also below the 92.7% reported in HyperMem’s original paper. The more explanatory result is therefore the improvement in connecting facts and temporal evidence, rather than a claim that every memory capability leads based on aggregate scores alone. Figure 8: Gains vary across question types; the document experiment uses M1 with AutoDream disabled Document question answering uses M1. SHDocQA and MHDocQA each evaluate 100 questions, with accuracy rates of 83.0% and 71.0%, respectively, 31.0 and 22.0 percentage points above HyperMem. On the 71 DetectiveQA questions, the result is 64.8%, below MemoryBank’s 67.6%. Precise facts and associative retrieval show clear benefits, while reasoning questions require further separation of retrieval omissions from reasoning failures after the evidence has been obtained. The user-side experiments use gpt-4.1-mini for answering and extraction. MIRIX uses text-embedding-3-small, while external baselines retain their respective default embeddings; session questions are scored by majority vote from gpt-4o-mini. These results come from a single configured run and reflect the entire pipeline. The paper’s ablations do not independently isolate the respective contributions of multifact frames, source readback, event identifiers, and answer verification. MemoryCD advances the question to personalized generation: for summaries of cross-domain reviews from 100 Home_and_Kitchen users, MIRIX-2.0 has the highest score on all four text metrics. What is measured here is similarity to reference text. The configuration record does not explicitly identify M0, M1, or M2, nor does it directly measure user satisfaction, so it is better suited to supporting the use case that memory can support cross-domain generation. 5.2 Skill Transfer The procedural-memory experiments first learn from training tasks, then freeze memory and evaluate on unseen tasks. ALFWorld uses 39 training tasks and 134 test episodes, while WebShop uses 40 training tasks and 140 test instructions. The same comparison uses identical actor prompts and models, and reports the means and standard deviations of three independent runs. This makes the experiment focus on whether established procedures can transfer, rather than allowing answers to continue accumulating during testing. Figure 9: Procedural memory achieves the highest observed mean on unseen tasks, while showing run-to-run variation On the paper’s GPT-6 Luna, ALFWorld’s success rate rises from 69.4% without memory to 76.4% with MIRIX-2.0, while AWM reaches 73.6%. In WebShop, the rate rises from 31.9% to 37.4%, while AWM reaches 36.0%. MIRIX’s standard deviation on ALFWorld is 6.2 percentage points, and it reaches 14.0 points on the other backbone, GPT-5.6 Luna. The three runs produce a relatively high mean while leaving room for further experiments to confirm stability. The task-level results deserve closer attention. On GPT-6 Luna’s ALFWorld, Clean and Heat improve significantly relative to no memory, while Pick, Look, and Pick Two decline. Existing procedures are most valuable when they fill in missing action structure; when the model can already handle the current task, additional procedures may also interfere. Figure 10: Skills have different effects on different tasks, and some categories fall below the no-memory result The GPT-5.2 results on MetaClaw-Bench provide another use case: the average question score rises from 44.0% with MetaClaw Skills to 49.2%, while the completion rate for file-inspection questions rises from 17.1% to 27.7%. The former covers all 346 questions, whereas the latter covers only the 224 file-inspection questions; these baselines are values reported in the original papers. Together, the two metrics suggest directions for further testing of skill reuse, while their respective denominators and comparison conditions need to be preserved. 5.3 Shifting Work Earlier The association index moves some work to the write stage. Across the three document tasks, M1’s ingestion time is 3.8–4.0 times that of M0, and its answers are also slower. However, M1 answers faster than HyperMem on five timing benchmarks—for example, SHDocQA takes 5.45 seconds per question versus 10.55 seconds—but this advantage comes with the upfront cost of building document graphs. Figure 11: Graph construction increases document ingestion cost, while M1 answers faster than HyperMem The two parts of the figure should be read together: the left side shows ingestion time in hours on a logarithmic scale, while the right side shows the average number of seconds needed to retrieve and answer each question. All timing measurements disable AutoDream and therefore correspond only to M1. The data do not cover the organization calls, total ingestion cost, or answer cost of the full M2 configuration. Skills likewise carry both learning and usage costs. In WebShop, MIRIX uses the fewest training and test tokens among the memory methods while achieving the highest average success rate; test tokens still increase compared with no memory. In ALFWorld, the average number of action steps in successful episodes decreases, but MIRIX has the highest total number of test tokens among the memory methods. Fewer actions may come with longer context per step, so whether this is worthwhile depends on task completion and subsequent reuse rather than step count alone. 6. Summary Returning to the original question: long-term memory must both preserve the past and improve future behavior. How can it avoid becoming increasingly distorted as it is organized and increasingly difficult to use as it accumulates? MIRIX-2.0’s answer is to draw clear boundaries between responsibilities. Committed records serve as authoritative evidence; multifact relations and event identities organize associations; query budgets control how much evidence is expanded; and AutoDream separately organizes identity entry points and action skills. The most valuable lesson in this approach is to separate observations that must be preserved continuously from indexes and skills that are allowed to change. Language models can propose extraction, same-entity judgments, and skill modifications, while deterministic code controls source binding, candidate scope, event identity, write order, and computational limits. The system therefore gains both a clear preservation boundary and a derived layer that can continue to improve. The next engineering questions are equally concrete: how many key details do the writers preserve, how accurate is entity merging, which skills genuinely help subsequent tasks, and how does the complete system amortize graph-building and organization costs as history grows longer? For personal assistants, viewing, correcting, deleting, and isolating user data also need to be incorporated into the deployment design. Continued validation along these lines is needed to advance the paper’s layered mechanism into a long-term dependable memory service. References • MIRIX-2.0: Multi-Agent Memory System: https://arxiv.org/abs/2611.12068

MIRIX Memory 元学习 Hypergraph AutoDream