cs.AISep 26, 2026
SaveAgentsensus: Consensus-Compressed Shared Memory for Multi-Agent Story Worlds
Organizations: University of Nebraska–Lincoln
Abstract
A agentic story world is a dynamic system simulating who learned what, when, and from whom -- yet the standard design gives each character a private memory stream. A shared event is therefore stored once per witness, large duplication will be incurred in terms of storage. We present Agentsensus, a story-world simulation framework in which there is an unified long-term memory. Records of the same event merge into one owned by all its witnesses, and semantically relevant memory records are linked. We evaluate on four worlds -- two classical Chinese novels, Hamlet, and a real-world conflict timeline -- run for 40 to 80 rounds against three per-character memory designs under an equal-granularity protocol. Agentsensus writes 22-44% fewer entries than the closest baseline and is the only design whose memory becomes shared (14-28% of records held by more than one character, some by 10) and linked (94-99%), at judged simulation quality indistinguishable or even better than the baselines. An ablation attributes this to the merge itself: disabling it multiplies the store by 3.1x and takes sharing to exactly zero. Sharing also compounds with the horizon rather than saturating early, rising 6% to 9% to 14% as one world is re-run at 10, 20 and 40 rounds.
Figures & tables
Figure 1: The framework shared by all four backends. A source text is sedimented into owner-tagged memory items that seed one long-term memory store; a round-barrier kernel then steps every awake agent, and the actions they take deposit into and retrieve from that store through short-term memory — a goal stack, a status register, and a cache of recent action–result pairs. The society panel names the six actions that act on the world and on other agents. Only the long-term memory box differs between the four systems compared in this paper.
Figure 2: The long-term memory structure of the four backends: private streams scored by recency, importance and relevance, over which reflections are distilled and themselves become evidence for further reflections; a two-tier graph whose insights are distilled from the interaction rows they point back to; a single store partitioned by access control, where sharing grants a second reader rather than merging the rows; and the consensus store, where equivalent deposits merge into one record whose owner set is the union of its witnesses, with affiliation edges joining statements split from the same deposit.
| Backend | Sediment | Sim | Total |
| Three Kingdoms — fiction, 80 rounds, 6,052 source events | |||
| Consensus | 6,051 | 974 | 7,025 |
| Gen. Agents | 20,896 | 1,251 | 22,147 |
| G-Memory | 27,156 | 1,434 | 28,590 |
| Collaborative | 19,856 | 1,455 | 21,311 |
| Red Chamber — fiction, 80 rounds, 6,506 source events | |||
Table 1: Store size by backend and world. “Sediment” is the store at round 0, produced by each mechanism’s own ingest of the identical source events — one row per event under consensus, one row per witness under every baseline, plus each backend’s own by-products (Generative-Agents reflections, G-Memory insight nodes). “Sim” is what the simulation then added and is the comparison the fairness protocol licenses: all four backends atomize identically, so entry counts are comparable by construction. Structure — what fraction of the sim entries becomes shared and linked — is Table 2 .
| World | Sim entries | Shared (multi-owner) | Linked (affiliated) |
|---|---|---|---|
| Three Kingdoms (80r) | 974 | 185 (19%) | 945 (97%) |
| Red Chamber (80r) | 438 | 105 (24%) | 412 (94%) |
| Russia–Ukraine (40r) | 814 | 114 (14%) | 806 (99%) |
| Hamlet (40r) | 130 | 36 (28%) | 127 (98%) |
Table 2: Sharing and linking across the four worlds, consensus store. Every baseline is at exactly 0% on both columns in every world — by construction, since no baseline merges records across agents or links atoms — so they are stated here rather than tabulated.
Figure 3: Memory growth, Three Kingdoms, 80 rounds. Left: cumulative simulation-written entries per round for all four backends, under the same accounting as Table 1 — the consensus store is below every baseline from the opening rounds and the gap widens, because each merge removes an entry that would otherwise have been written and re-recorded by later witnesses. Right: per-agent owned memories in the consensus run, busiest six agents labelled. The other three worlds repeat the pattern (Appendix H ).
Figure 4: Memory-operation latency, Three Kingdoms: mean seconds per remember (left) and recall (right), 5-round bins. Writes cost tens of seconds wherever the mechanism calls a model on the write path — consensus for the equivalence judge, Generative-Agents for importance and reflection — while reads stay under about a second for every backend, because retrieval calls no model. The profile repeats in the other three worlds (Appendix H ).
Figure 5: Continuation quality — Three Kingdoms, 80 rounds: grounding, trajectory, narrative and goal pursuit, bars are means over three scorings with std whiskers. The first three panels are scored from the rendered screenplay, the fourth from the event log. The other three worlds, and the full table of means, are in Appendix I .
Figure 6: The three layers of one run — Three Kingdoms, 80 rounds. (a) Interaction graph: each node an agent, sized by degree, coloured by conversational community; the court around the emperor, the Liu Bei camp and the Jiangdong envoys separate without being told apart. (b) Memory-affiliation graph: each node one consensus memory, edges are affiliation, colour is cluster; three memories from different clusters are circled with their text and owner set — a cluster is a plotline, and its memories’ owner sets name the agents living it. (c) Shared-ownership heatmap: cell counts memories owned by both agents (log scale); its block structure is (a)’s communities seen from the store. (d) A subset drawn in full: five conversing agents (blue edges, weighted), the memories they own (dark lines and larger nodes where a memory has several owners), affiliation arcs among those memories (purple), and three memories labelled with text and owners — one record, several owners, is what consensus compression produces and what every baseline lacks. The other worlds are in Appendix J ; in the HTML version these figures are interactive.
Figure 7: The three pairwise relations, Three Kingdoms. Each panel contrasts pairs that have the property with pairs that do not: agents who converse against agents who do not, linked memories against unlinked ones, and cross-agent memory links against the rest.
| Merge | Cache | Store | Sim | Shared | Linked | Grnd | Traj | Narr | Goal |
|---|---|---|---|---|---|---|---|---|---|
| on | fifo (published) | 6,590 | 539 | 19% | 93% | 0.92 | 0.89 | 4.17 | 0.97 |
| on | relevance | 6,674 | 623 | 21% | 96% | 0.93 | 0.83 | 4.33 | 0.99 |
| on | hybrid | 6,670 | 619 | 18% | 96% | 0.94 | 0.90 | 4.25 | 0.98 |
| off | fifo | 20,669 | 813 | 0% | 94% | 0.92 | 0.82 | 4.00 | 0.96 |
Table 3: One factor at a time from the published configuration, consensus on Three Kingdoms, 40 rounds. All four quality metrics here read the event log, so they compare across cells rather than against Table 9 .
Appendix figures & tables18 assets
Supplementary material from the paper’s appendix.
Appendix
| recall | affiliated | keep-shorter | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| World | rows | hits+exp. | amp. | q in hits | q in exp. | 1 4+ own. | seen | chars | name | principal | |
| Three Kingdoms | 7,025 | 51 | 5+27 | 6.6 | 25% | 9% | 8 13 | 60/447 | 4 | 30% | 15% |
| Red Chamber | 6,944 | 19 | 5+35 | 8.7 | 32% | 9% | 4 19 | 43/267 | 5 | 37% | 7% |
| Russia–Ukraine | 2,347 | 41 | 5+1 | 1.6 | 10% | 0% | 7 1 | 37/314 | 20 | 22% | 14% |
| Hamlet | 1,265 | 3 | 5+40 | 8.8 | 0% | 0% | 2 26 | 7/74 | 9 | 29% | 14% |
Table 4: What the mechanism costs, per world. recall : number of calls, median semantic hits plus rows added by one-hop expansion, the resulting amplification, and how often the query string appears in each group. affiliated : median siblings carried by a row at one owner versus four or more. keep-shorter : merges whose discarded side the log preserves, over all merges; the median characters discarded; the share of those merges whose surviving text loses a named character; and the share where the lost name is the party that ordered the act.
| Action | Consensus | Gen. Agents | G-Memory | Collaborative |
| Three Kingdoms | ||||
| say | 66 | 50 | 48 | 61 |
| read_thread | 123 | 110 | 83 | 121 |
| observe | 28 | 34 | 31 | 37 |
| move | 2 | 1 | 4 | 1 |
| act_on | 11 | 14 | 23 | 22 |
Table 5: Actions called per backend, by world, over each world’s final stage. Counted from the event logs. Discretionary memory-management actions are omitted because every count is zero.
| World | Rounds | Scenes | Beats | Speakers | Places | English | Source lang. |
|---|---|---|---|---|---|---|---|
| Three Kingdoms | 0–79 | 54 | 312 | 33 | 24 | 206k | 62k (zh) |
| Red Chamber | 0–79 | 62 | 390 | 34 | 26 | 98k | 30k (zh) |
| Russia–Ukraine | 0–39 | 48 | 334 | 47 | 20 | 313k | – (en) |
| Hamlet | 0–39 | 12 | 117 | 16 | 6 | 49k | – (en) |
Table 6: The four screenplays. “Beats” are the dramatizable events after deduplication; “scenes” the groupings by place and stretch of time that each become one render call. Length is the rendered markdown; the Chinese worlds carry two renderings of the same beats.
Figure 8: Hamlet, 40 rounds: 117 beats across six places cut into twelve scenes. The busy hall at Elsinore becomes two scenes rather than one because the twenty-round cap divides it; the single action in Norway is a scene of its own.
Figure 9: Three Kingdoms, 80 rounds: the pattern of a campaign narrative. Four places carry the war in long unbroken bands while a dozen others are visited once and become one-cell scenes.
Figure 10: Red Chamber, 80 rounds: the densest world in the set. The household keeps returning to the same handful of courtyards, so four of them split into three or four scenes each.
Figure 11: Russia–Ukraine, 40 rounds: institutional actors work from fixed seats, so activity concentrates in a few capitals and headquarters. The hatched cells — several institutions acting in one place and round — are where the many-witness merges of § 5.1 come from.
| World | Registry | Events | Calls | Tokens | Wall | Active | Archived | Envs | Carriers | Warn |
|---|---|---|---|---|---|---|---|---|---|---|
| Three Kingdoms | 399 | 6,052 | 764 | 9.2M | 124 min | 33 | 38 | 115 | 5 | 235 |
| Red Chamber | 152 | 6,506 | 848 | 8.5M | 129 min | 34 | 3 | 88 | 9 | 222 |
| Russia–Ukraine | 170 | 1,533 | 1,373 | 12.3M | 69 min | 47 | 14 | 71 | 0 | 61 |
| Hamlet | 22 | 1,135 | 212 | 0.9M | 31 min | 16 | 2 | 8 | 3 | 10 |
Table 7: Sedimentation cost and cast composition. Wall-clock and token figures are for the sedimentation pass only, on the consensus store; the per-backend ingest of the same events is additional. “Warn” counts extraction records the pipeline flagged for review.
| World | Sim entries | Shared | 3+ | Max | Linked | Expanded | Linked/call |
|---|---|---|---|---|---|---|---|
| Three Kingdoms (80r) | 974 | 185 (19%) | 38 | 6 | 97% | 51/51 | 28 |
| Red Chamber (80r) | 438 | 105 (24%) | 24 | 6 | 94% | 19/19 | 38 |
| Russia–Ukraine (40r) | 814 | 114 (14%) | 23 | 10 | 99% | 23/41 | 5 |
| Hamlet (40r) | 130 | 36 (28%) | 2 | 3 | 98% | 3/3 | 39 |
Table 8: Consensus structure per world at the final stage. “3+” counts entries with three or more witnesses, “Max” is the deepest merge, “Expanded” the fraction of recalls that returned at least one memory reached along an affiliated edge, and “Linked/call” the mean number of such memories per expanding call.
Figure 12: Red Chamber, 80 rounds: growth (top) and per-call latency (bottom), read as Figures 3 and 4 .
Figure 13: Russia–Ukraine, 40 rounds: growth and per-call latency.
Figure 14: Hamlet, 40 rounds: growth and per-call latency.
Figure 15: What the merge costs per call: mean seconds per remember (left) and recall (right), 5-round bins. The equivalence judge is a second LLM call on the write path, so merging roughly doubles the cost of a deposit and the gap widens as the store grows; reads are unaffected, since owner-scoped retrieval and one-hop expansion call no model.
Figure 16: Continuation quality: Red Chamber (top), Russia–Ukraine (middle), Hamlet (bottom), read as Figure 5 .
| World | Backend | Grounding | Trajectory | Narrative | Goal |
|---|---|---|---|---|---|
| Three Kingdoms | Consensus | 0.88 0.02 | 0.87 0.02 | 4.08 0.12 | 0.98 0.00 |
| Gen. Agents | 0.93 0.01 | 0.82 0.02 | 4.00 0.00 | 0.99 0.00 | |
| G-Memory | 0.88 0.01 | 0.87 0.01 | 3.83 0.12 | 0.99 0.00 | |
| Collaborative | 0.89 0.02 | 0.89 0.01 | 3.92 0.12 | 0.97 0.01 | |
| Red Chamber | Consensus | 0.92 0.02 | 0.86 0.02 | 4.00 0.00 | 0.98 0.01 |
| Gen. Agents | 0.88 0.04 | 0.87 0.02 | 4.25 0.00 | 0.98 0.01 |
Table 9: Continuation quality, means over three scorings, all worlds and backends. No backend leads everywhere and no world has one leader on all four metrics.
Figure 17: Red Chamber, read as Figure 6 .
Figure 18: Russia–Ukraine, same panels.
Figure 19: Hamlet, same panels.
More in this topic
Agent Memory Management
Agent memory management studies how AI agents store, update, retrieve, compress, and forget information across long-running tasks.
See all 449 papers